IARPA AGILE Proposers Day

Show video

Good afternoon and welcome to the IARPA AGILE  virtual Proposers' Day my name is Bill Harrod   I am the Program Manager for the AGILE program we  are here to present a very exciting program it is   called the Advanced Graphical Intelligence Logical  Computing Environment program which we refer to as   AGILE yes the word computing is not represented in  the acronym for the program title that's because   the program is focused on developing new computer  architectures for data intensive applications not   compute intensive applications uh besides the  'C' would mess up my acronym thank you for your   interest in learning more about the IARPA and the  IARPA AGILE program and doing business with IARPA   a few notes before we start the presentations  during the day please submit your questions   and comments you'll notice a Q&A button on the  bottom right hand corner of the WebEx screen   use your this button to enter questions and  comments other participants will not view your   questions or comments we want to receive your  questions and comments the topics could include   IARPA the technical presentation doing business  with IARPA or the technical section from the BAA   later in the day we will select some questions  and provide answers the WebEx audio and video   for all participants will be disabled a recording  of the entire day will be posted on the AGILE BAA   website this presentation provides information  concerning IARPA and the AGILE program the   presentation will assist you with your plans  for a proposal submission to the eventual   AGILE BAA the Proposers' Day does not constitute  a solicitation for proposals white papers or   abstracts the BAA will be improved based  upon the feedback that we receive during   and after the proposal state so I encourage you  to provide us with your questions and comments   your questions and comments will  influence the development of the BAA The goals of this meeting are to provide  information concerning IARPA and the AGILE program   hopefully I will also hopefully  it will also facilitate teaming   the successful execution of the AGILE program  will require teams that have a diverse set of RD   experience and talent the teaming form posted  on sam.gov will enable proposers to find other   groups that have capabilities that are required  to establish a strong team during the meeting   there will be two breaks the first will last  15 minutes and the second will last 30 minutes   during these breaks I encourage the participants  to submit their questions and comments we must   receive your input before 2:15pm we will provide  answers to selected questions starting at 2:30.   We will read all questions and comments  they will have an impact on the eventual BAA   this is the agenda for the Proposers' Day meeting  our next presentation will be by Pedro who is the   Director of the Office of Collections at  IARPA he will provide an overview of IARPA   after Pedro's talk I will present an overview  of the adjust program this will be followed   by a 15-minute break during this break you  should formulate and submit your questions and   comments at three at 1:35 Pedro who is going to  pretend he's a contract officer will talk about   doing business with IARPA after Pedro's talk we'll  have a 30 minute break during this time you should   submit your final questions and comments please  finish before 2:15 pm at 2:30 I will provide   answers to selected questions thank you for  listening and now we will start with Pedro's talk   so Bill thank you so much for that good  afternoon to everybody I am Pedro Espina   I am the Director of the Office of Collections at  IARPA Bill will you please be so kind to continue   to move the slides forward for me absolutely  thank you okay so let's go to the next slide So as many of you might be aware IARPA  is part of the United States Intelligence   Community the United States Intelligence  Community is formed by 16 agencies and at   the head of it is the Office of the  Director of National intelligence   IARPA is part of the Office of the Director  of National Intelligence our function is that   to serve as technical uh know-how to the entire  Intelligence Community till the next slide please So IARPA operates this way we envision and lead  high-risk high-payoff research that delivers   innovative technology for future overwhelming  intelligence advantage our problems typically   are very complex and their nature is typically  multi-disciplinary we emphasize technical access   and technical truth okay so the truth guides us in  everything that we do at IARPA next slide please We try to use a full open competition in our to  solve our problems to the greatest possible extent   why do I say that because sometimes some of  our problems that we tackle are classified   and when they are classified we can only use  full open competition among people that have   their appropriate clearances to help us with those  problems in addition to that we have a category of   world-class program managers who typically rotate  in and out of IARPA into academia industry or   government every three to five years nobody from  the Director down in IARPA is a permanent employee When we execute a research program at IARPA  we have very clear goals that can be measured   uh and we have those goals are very ambitious but  they are credible we don't try we're not trying   to do we're not dealing in science fiction  we're trying to deal with things that are   within the realm of the possible we employ  independent and readers testing and evaluation   we form government teams that help us come up with  the testing and evaluation of the performance of   those people participating in any one of our  programs we involve the intelligence community   throughout the process our partners from IC  agencies come in at the beginning to tell us   what their most challenging problems are  they observe how we progress throughout   the entire process of developing the program  and then they accept our research as input   into their operations okay our programs like  I said before last typically between three   and five years we have some little programs  that we call seedlings that typically last   between nine and twelve months but the main  programs that ones like AGILE today will last   between three and five years we encourage our  performers to publish in peer review papers   why because that meets the criteria of good  scientific work okay in some instances we   cannot do that because some instances some of  the work that is being produced is classified   and in those cases we still try to publish  it in peer review papers that are classified   and we try to transition every capability that  we develop into intelligence community partners   in other instances sometimes we transition  our capabilities to people that are in other   government agencies still click the next slide  so the IARPA research is very very broad okay   it includes programs in mathematics programs in  physics chemistry biology neuroscience linguistics   political science cognitive philosophy and much  more okay about since the since IARPA was stood   up about 70 percent of all the work that we  have done has actually transitioned into other   government agencies the that work has produced in  excess of 3,000 peer-reviewed journal articles and   some of the people participating  in IARPA programs have actually   been awarded with very high accolades among them  the Nobel prize in physics for quantum computing   the McArthur fellowship Nobel prize and  others we are connected inside government   all the way from the White House to the  bench scientists in a national laboratory   at the level of the White House we participate  in the national science and technology   council activities for example we have been  participating in the White House Brain Initiative   in the national strategic computer initiative  in the select committee artificial intelligence   on the subcommittee on quantum information  science and the subcommittee on economic   and security implications for quantum  science so we're plugged everywhere   where science is needed by the government  of the United States next slide please So how do you work with us so this afternoon our  contracting officer had an unfortunate situation   today so I will have to be the person talking to  you about how to work with us from the contracting   officer perspective but for now let me tell you  this the principal vehicle for you to contact   us is our website IARPA.com you have a phone  number in this slide that you can call us   and your route your phone will be your phone call  will be routed to the appropriate person within   the structure we also have a mail dropbox called  info@iarpa.gov where you can send your questions  

and somebody will typically answer to you between  24 and 48 hours later um we have a number of ways   to engage with us um RFI uh actually if you go  into uh beta.gov you will see that presently   we have some RFIs out and those RFI typically  lead to workshops like I said before we have   these little programs that we call seedlings that  last between 9 and 12 months and typically we use   those to turn uh doubt into disbelief as  we are actually handling a new concept   we sometimes put prices prize challenges out on  the street to see what comes back and we built   very large research programs like the one we're  talking today agile okay and with that I think   that next bill will be actually talking about  the bones of AGILE as a technical program okay   and I will be back with all of you around 1:35  so you can actually see how ignorant I am about   contracting law okay thank you so much for  your time if you have any problems please   type into the Q&A section telling us that you're  having problems that you're gonna hear or see   us or whatever and we will try to address  your technical needs okay Bill back to you Hi thank you Pedro again this is uh Bill Harrod  I am the Program Manager for the AGILE program   I'm here today to present the  details concerning the program   we anticipate the BAA will  be released in early January The Program's purpose is to develop new computer  architectures that are efficient and scalable   when converting massive and diverse data sets  into actionable information in a reasonable   time frame computers that are available today and  within the next few years are truly awe-inspiring   however today's computer systems are not designed  to provide optimal performance for future data   analytics problems these computers will limit  the size and complexity of the data analytics   problems that can be tackled and make poor use  of the resources we need to have a significant   departure from the previous 60 plus years of  computer development to resolve these problems   we need a system level design approach  driven by real applications we cannot use   a band-aid approach to improve the components  independently we need to start at the system level   this is my agenda for today uh topics that I will  be covering uh we'll start with what is AGILE   and why are we doing this program we'll discuss  technical challenges and there are probably many   more than what I have presented or will present  we'll also discuss a few research strategies   to resolve the technical approaches we'll talk  about the design process and the expectations   we'll talk about how success is  measured and what is the schedule But first let's take a look  at what is AGILE really about   analyzing massive data analytics problems is  an important but extremely challenging effort   the digitization of society science and security  has resulted in massive data sets that are complex   and rapidly changing there are extremely large  sources of data that include social networks   public data sources sensor data and many more  sources of data the data sets can include billions   of elements the data could be streaming to a  computer or found in data lakes or warehouses   the various relationships and events found  within the datasets could be constantly changing   streaming data will require new algorithms  or implementations of existing algorithms   the items outlined in red in this diagram  show where the AGILE program will focus   the object is to generate actionable  knowledge in on 10x more data   than today's capabilities at a rate of 10  to 100 times faster than today's computers IARPA is seeking new revolutionary computer  architectures that are solving the emerging   data analytics problem AGILE is envisioned to be a  three-year program beginning of approximately June   2021 and extending through June 2024 performers  will be provided with four workflows and data sets   there are target metrics for each workflow  and benchmark codes performers will develop   RTL designs for their proposed architecture  including a runtime system designs will be   evaluated using an AGILE enhanced modeling  simulation tool set called assist that is   based on the Sandia National Laboratory software  the structural simulation toolkit or SST SST is an   open source toolkit there are numerous tutorials  and documentation available on the internet The data of interest can't be transformed into  actionable knowledge in a reasonable time frame   unfortunately often events and relationships  are discovered after they have occurred   the data sets are increasingly sparse random and  heterogeneous the analytics are characterized by   having minimal data locality poor data reuse fine  grain data movement and data driven parallelism   the fundamental problem is that today's  computers solve yesterday's problems   they are not efficient or scalable  for emerging data intensive problems   we have a data movement and storage problem not  a compute problem we need a new system level   intelligent mechanisms for moving accessing and  storing large random time-varying data streams   and structures designs need to be driven  by end-to-end workflows not simple kernels Let's talk about how data is transformed in  knowledge often data is represented by a graph a   vertex in the graph represents an entity and the  edges represent relationships between entities   these graphs tend to be very large and have very  few relationships between the entities or vertices   often the graph is represented by a matrix thus  the resulting matrix tends to be very large and   sparse algorithms come from a variety of classes  ranging from graph analytics to machine learning   to linear algebra for streaming problems the data  is constantly changing and randomly distributed   across the system we are drowning in data and not  able to analyze it before its full values can be   exceeded data types vary from  measurements to documents   many algorithms incorporate reasoning  processes where there are issues concerning   the heterogeneous integration of graphs and  databases thus the vertices and edges have   a contextual locality property as opposed  to the ordinary spatial or temporal locality   the surrounding data provides meaning or context  will be found within a diverse set of data sets   to fully understand the context requires  accessing from a variety of sources   this is the fundamental problem of the AGILE data  that causes the greatest computational challenges The AGILE challenge problem is dealing with  streaming data efficient and timely execution   of the analytics will allow us to transition from  a forensic analysis environment to a predictive   environment having the capability to analyze  why an event occurred is extremely important   but what is what we really want is the ability  to predict that event could occur streaming   data analytics provides a continuous flow  of information from updates on the data to   um uh queries and analysis of streaming graphs  helps us understand extreme scale and dynamic   real-life interactions and helps predict future  events this has motivated the development of   streaming algorithms to incrementally update  the graph analysis moving away from performing   nightly updates of graph properties to providing  these updates in minutes or seconds the bottom   line is that systems need to be designed for  rapidly changing sparse data and processing flow Now we'll examine some of the challenges  associated with the AGILE computational problems   today the results are required in near real time  up to hours identifying event after it occurred   doesn't have the same impact as knowing that  an event might occur predicting the future   streaming data causes unpredictable changes  to store data often the computations involve   extremely fine-grained data movement where  the data is distributed across the computer   data computation tasks are typically determined  by the data and the streaming queries   but these tasks have extremely poor locality and  data reviews many graph algorithms can be recast   as sparse linear algebra operations recently  researchers have been furthering this approach   to a graph analytics for further background  information I encouraged you to examine the   book graph analytics in the language of linear  algebra written by Jeremy Kepner and John Gilbert Next we'll examine a few examples that demonstrate  the low efficiency and scalability properties   of several graph analytics computations  the sparse matrix times a vector called   SpMV is a standard algorithm that can be  used in many graph analytics implementations   Brian Page from University of Notre Dame optimized  a distributed version of SpMV and used the   software to compute the operation for an extensive  collection of sparse matrices this graph shown on   this slide shows the performance of the results  for the densest matrix and the sparse matrix   as you can see from this chart the distributed  version of SpMV doesn't scale in fact for the   Sparse matrix the computation probably should  have been simply computed on a single CPU   for the AGILE data of interest where the  data is randomly distributed across the   system memory space the performance results will  be much worse this chart shows the performance   of the industry standard 8 HPCG benchmark these  results were reported recently in November of   2020. You'll find much more information about  this benchmark on the HPCG benchmark suite the   benchmark uses an internet algorithm for solving  large sparse linear systems the main components of   the algorithm are SpMV that's the first matrix  times a vector global sum and synchronization   the fastest system as determined by the benchmark  is also the world's newest and fastest computer   computer however the HPCG benchmark for  this system is only 3.6 percent efficient   yet the it is 82 efficient for the  HPL benchmark which solves a dense   linear system and provides you with  the results for the top 500 benchmark Finally this chart shows the performance  of the graph 500 benchmark for a specific   multi-CPU computer these results were  generated by Peter Kogi at Notre Dame   the graph 500 is a benchmark standard benchmark  for data intensive applications there are two   versions of the benchmark the results shown  in this graph are for the breadth first search   version as you can see from the chart for  the grass 500 results it takes approximately   100 CPUs to achieve the same  performance as a single CPU   the chart also shows the performance for HPCG on  the same system it takes over a thousand CPUs to   achieve the same performance for one or several  CPUs the traditional HPCG system is not capable   of providing scalable performance results for  this class of problems or the AGILE data problems So as previously said today the HPCG systems  that are available and emerging in the next   couple years are truly impressive computers  so what is the problem with today's computers   fundamentally the problem is that the  original designs were driven by 3D   multi-physics simulations the fundamental  computation infrastructure has remained the same   over several generations of designs vendors  are interested in improving the performance   of a computer by taking advantage of faster  CPUs or GPUs utilizing stack dram or nvram   and 3D packaging technology so this works  extremely well when you're focusing on a node   but the overall efficiency of the computer is not  optimized for the emerging data analytics problems Today's computers have three primary subsystems  these are communication memory and computation   the communication subsystem is the network that  interconnects memory and computational elements   for most computers there are  actually two separate networks   often they are designed to move blocks of  data not optimal for individual data items   a memory system provides mechanisms  for accessing and storing data   the memory architecture typically provides optimal  performance for cache line loads there is no local   processing and the CPU is actually what controls  all of the data movement finally the computation   subsystem provides mechanisms for executing and  controlling the floral path most are based on   a von Neumann architecture which appears to  be poorly suited for data intensive problems   today's computers are mostly defined by  commodity parts and interfaces for example   a computer could be based on nvidia GPUs  and CPU GDR or GDR memory and infiniband   vendors are constrained by the standardization  of the interfaces that determine the properties   of data movement operations and are resistant to  making major changes in the system architecture   industry has been focusing on improving CPU GPU  components and local memory accesses but not on   global data movement or processing random data  addresses or changing the computational model What we need to do is tear down the traditional   barriers or boundaries between communication  memory and compute elements in the design   we need to change the overall system  model yes break 60 plus years of tradition   it's important that the AGILE architectures  provide a highly productive environment   clearly this change in the computing paradigm will  impact the software productivity on our computer   the AGILE program is not seeking to develop a  special purpose computer that will result in   only one computer being built if the computer  is extremely challenging to program or execute   programs then its widespread usage will be very  limited the blurring of the component boundaries   will increase the complexity associated with the  development of software and its execution on a   computer however the runtime system will have a  significant and positive impact on this problem   thus a runtime system becomes the  fourth major subsystem in a computer   it is what will glue everything together it is how  high productivity will be enabled on the system Let's examine six technical challenges  associated with the data analytics problems   and as I said previously there are probably  several that I have not included in here   first is designing the processing  elements to execute efficiently   a broad spectrum a data driven task perhaps  not known until execution time data analytics   workloads can have an execution patterns  with poor spatial or temporal locality   and instruction streams with numerous branched  conditions that cannot be predicted ahead of time   next area is supporting fine grain irregular data  movement ranging from the processing elements to   the system level data irregularity will arise  in both data layout and compulsory data movement   to keep a data movement pathway fully  occupied which is required to get   the best back bandwidth out of a  communication system uh irregular Sorry skip online next we support fine grain data  fine grain irregular data movement ranging from   processing elements to system level to keep  the data pathways fully occupied there must   be enough data movement requests in progress  concurrently that will keep the channel full   and thereby hiding the latency algorithms will  exhibit data-driven parallelism at all levels   fine-grained to core screen graph processing  for irregular time varying structures has   parallelism that can only be discovered  in real time we need a new memory storage   architectures for random dynamic to regular  data structures AGILE workflows will have   non-regular structures whose number shape and  size are not known until runtime and may change   dynamically during runtime since the data  properties may be highly varied and unknown   until runtime the system and even processing  elements may need to have dynamic introspection The I/O system needs to be capable of ingesting  high velocity streams of data from multiple   external sources file semantics make it  challenging to ingest data from multiple   data sources into a file interface that has the  sequential semantics while still maintaining   global consistency across all clients system  security data integrity compliance services   need to support multiple applications  cooperating with in a same memory space   for modern high-performance computing systems  multi-tendency usually involves many different   users and applications from different access  credentials to share the common database So in summary we've touched upon several of  the challenges um associated with the emerging   graph analytics problems these range from a  fine-grained unpredictable date of movement   to various levels of extreme parallelism  to the need for computational elements   that are designed for data intensive  computations and security and compliance As I previously said what we have to do here  is break the 60 plus years of traditional   architectural design and these all started  back in 1945 with the von Neumann architecture The primary characteristics of a von  Neumann architecture are basically that   they are optimized for AOU utilization there's  a separation of compute and logic and memory   and there is a some sort of mechanism for  providing sequential instruction issue let's   examine some of these requirements and why they  necessitate a non-von Neumann architectural model   for systems designed with analytic problems of  the FPU or ALU isn't the most valuable resource   it's operations associated with data movement  in storage that are actually the most important   future systems could take advantage of  computing in memory or pre-processing the   data before data movement and thus the  need to move the data could be limited   latency hiding schemes for data analytics  requires asynchronous operations that   result in unpredictable instruction issue thus  breaking the sequential instruction capability   global namespace will enable many  features including moving compute to data There are more issues associated with  data intensive applications and many   examples strategies for addressing  them I am only presenting several of   them and the ones I presented aren't  the only ways to address these issues As you can see from the slide   this slide there are numerous research strategies  for overcoming the AGILE technical challenges   this slide provides a few of them this gives me  confidence that the AGILE program can succeed we   just have to be willing to tear down the barriers  first involves the development of the integrated   system level design again you have to start  from the top not with a box full of components   however I am neither advocating or discouraging  any of the strategies you find on this slide   I'm simply pointing out that there are possible  strategies other strategies could impact an AGILE   architecture the critical point is that your AGILE  proposal needs to propose a coherent research   strategy and why it resolves the technical  challenges it is certainly possible that this   presentation and the resulting BAA don't state  all technical challenges and research strategies   proposers should specify what they believe are  the technical challenges associated with emerging   data analytics problems and the  justification for those statements   and what research needs to be  performed to resolve these challenges Okay the AGILE program what is it all about  well the program objectives are first to   process 10x more data second is to achieve data to  acknowledge in 10 to 100 times faster than today's   capabilities and three achieve the AGILE target  metrics research teams will be funded to develop   new validated designs that can achieve the program  objectives and target metrics there will be   independent testing evaluation team that will  evaluate and validate the design the program   will have two phases each lasting 18  months to improve and mature the designs   during the first phase research teams  will develop a high level behavioral model   with a performance model for  the proposed architecture   during the second phase research teams will  develop an RTL model for their architecture the   metrics and and the testing evaluation process  will be discussed later in this presentation The AGILE program will develop cost-effective and  open designs this means it will not involve exotic   expensive technologies that cannot be realized  in silicon before 2020 before 2030. During the  

program performers will openly discuss their  designs the AGILE PM will include this information   in future public and private presentations the  designs will be developed using a co-design   process that is driven by the four workflows  benchmarks and of course the target metrics This is an important slide proposers need to  understand what should not be included in their   proposed effort for example the following  will not be funded the design of the power   distribution system a project where the design  methodology is not based on a co-design process   the implementation of IP that isn't  open or licensable by other parties   a system design that doesn't scale from a  single node to a larger multi-node system   also the program does not include the  development of the programming environment   however the runtime system and memory model  will clearly have an impact on programming   proposers should provide in their proposal a  vision for the anticipated programming environment   address the issues concerning how programmers  will develop applications for the new architecture   describe the methodology for converting a program  written in a high-level language into an optimal Executable okay as I said at the beginning  there are four workflows that will be provided   there will also be kernels associated with  the workflows and benchmark codes that are   available today performers will have access to  the source code for all of these applications   you'll also be provided with um the target metrics  and the BAA I list them here in the presentation   um I encourage everybody to pay close  attention to target metrics within IARPA uh a   performer's ability to achieve target metrics  is a very important measurement of success Okay so as I said there are four workflows  the first one is knowledge graphs   this involves groups relationships and interests  detection this includes system and event patterns   sequence data this means the identification and  clustering of data sequences and network cyber   which is the security of network systems for  each workflow kernels will be generated to   exercise one or more of the algorithms associated  with the metrics and finally there are three   industries standard benchmarks that will be  provided the first is breadth first search which   is graph 500 the same as graph 500 I should  say triangle counting and Jaccard coefficients The workflows shall drive the design  efforts this is accomplished by using   a co-design process that allows the evolution  of the design and implementation of workflows   so that the resulting architecture achieves the  best performance for all workflows the performance   results will be evaluated using the agile enhanced  modeling simulation toolkit called assist it is   based on the Sandia National Laboratory software  structural simulation toolkit SST thus clearly   based on my diagram the co-design process involves  three major components the workflows your design   and the validation process the chasm between  the demands of today's escalating data intensive   problems and the capabilities of yesterday's  computing systems is unbridgeable evolutionary   improvements are now providing diminishing returns  we need a fundamental architectural approach   a fundamentally new architectural approach for the  design of computers for data intensive problems   I encourage proposers to take advantage of this  program and develop architectures that are truly   revolutionary we are not interested in  a simple refresh of today's computers   resulting architecture needs to not just achieve  the target metric it needs to lead towards a new   class of highly productive computers tear down the  barriers by starting with a clean sheet of paper   head towards productivity your proposal should  discuss the impact of the design on productivity The funded research teams are expected to be  comprised of researchers and developers with   expertise in communication memory computation  runtime subsystems data analytics applications   and our focus on providing architectural  innovations for data-driven computations   the primary tasks are listed on this chart at  the most basic level though research teams are   developing system designs that can be demonstrated  to achieve the target metrics using assist This this chart shows the  different responsibilities   for the research team and  the test and evaluation team   the research teams are responsible for  developing the design the modules for a assist   and the models that are used with an assist that  are necessary to estimate the performance using   assist and tailoring the implementations of the  workflows kernels and benchmarks for their designs   the testing evaluation team will validate  the design and verify the performance   using assist they will evaluate the tailored  versions of the workflows kernels and benchmarks   while it is certainly true that the research  teams are responsible for utilizing the co-design   process the testing evaluation team will monitor  its usage and correctness in the design process   so let's take a look at a runtime system because  this is a very important part of this program   there is more to this than just doing silicon  design the design of a runtime is an abstraction   of the computing system software structure  and operation for a specific system model   it provides a conceptual framework for  the co-design of the computer including   the architecture programming interfaces  and system software the attributes of   the main attributes of the data analytic  challenges are extreme parallelism asynchrony   self-discovered parallelism adaptive  management and global name space An important aspect of the AGILE program is that  performers are developing a holistic system model   it is not as I said several  times a collection of parts   performers will need to start by  establishing a system level performance model   this requires consideration of the  performance factors for the model   in this example four factors are presented on  this slide these are starvation latency overhead   and weighting these factors are all related to the  utilization of system resources and the execution   of parallel work across the system proposers  should describe their own system level performance   model including the performance factors which  may not be the same as listed on the slide   describe how the performance model  will drive the design decision Many of the design drivers that impact AGILE's   performance models performance factors are not  uniquely caused by the data analytics problems   so one should keep that in mind let's talk  about some examples of how data analytics   affects slow performance factors starvation of  a computing resource will occur when there isn't   sufficient parallelism due to the limited number  of tasks or threads in a data analytics algorithm   increased latency will delay an operation  when a data atlantic operation is gathering   small data packets from several memory locations  and there are insufficient paths to the data   overhead which is wasted time is the time spent  setting up data analytics transfers for data   involving small packets waiting for a  remote memory occurs when the memory   remote memory has a large number of outstanding  small memory requests proposers need to specify   how data analytics problems impact the performance  of the performance factors in the model As I said before there are four workflows these  are all representatives of problems of interest   in the intelligence community again  to repeat myself there are four   graph knowledge graphs detection sequence  data and network but now we're going to talk   about these in a little bit more detail and  I'll also discuss the target metric but the   BAA will provide more details than what I'm able  to fit into this presentation knowledge graph   workflows discover new entities and relationships  traceable evidence paths complex events anomalous   patterns and complex system models the goals for  this workflow are to increase the data ingestion   rate by a factor of 100 reduce the time required  to both entity classification relationship   discovery by a factor of 50 reduce transformation  of data to knowledge from days to minutes Detection workflow identifies events event  requirements and temporal evolution precursors   and key event drivers and influences the AGILE  goals for this workflow are to increase the   size of the graph that can be processed by  a factor of a hundred I'm sorry a thousand   increase the number of data streams that can be  processed simultaneously by a factor of three   increase the capability from single event  detection to multi-detect invention an enable   incremental detection analysis incorporating  new data without the need to re-compute   both of which cannot be done today  because of computational limitations Sequence data workflow seeks to extract common  structures inheritance and functionality from   massive temporal sequences of data the AGILE goals  are to increase the number of sequence analyzed   per unit time by a factor of a thousand reduce  time to results from months to hours or days   and estimate uncertainties or priority  events something that is not done today Finally there is the network cyber workflow  it models complex interconnected networks   and their defense examples include power  grid comms and control social networks etc.   the AGILE goals are to reduce the time to identify  key influential energies by a factor of 60   identify and prioritize anomalous activities and  react to attacks in near real time so finally   there are the graph benchmarks um this chart shows  you these the three benchmarks that we're using   there's brett's first search triangle counting  and Jaccard coefficients I've also listed the   target metrics for two different  problem sizes and what must be reported So a lot of information um has been  provided we discussed the program objectives   benchmarks and workflows I want to give  everybody insight into how this all fits   into the overall effort from now and  throughout the life of the program The objectives benchmarks and workflows play  a critical role before and during the program   proposers must analyze the benchmarks and  workflows in the proposal and describe   how their design will impact these codes and  provide an estimated performance for the metric   during the program research teams will  utilize the benchmarks and workflows in   a design co-design process that will result  in optimal performance for these codes   throughout the program research teams will  report on the estimated performance of the   benchmarks and workflows for their  designs that are under development So the deliverables there are basically um a few  this chart provides the primary deliverables for   each phase if you look in the technical section  for the BAA you will I'm sure have seen a much   more extensive list of what the deliverables are  um but these are the primary ones of interest   the big ones of course and basically um it comes  down to the designs with supporting documentation   test plans and the modification  of the application software Also the performers need to provide an  application framework that will enable   the the construction of the executables  for the workflows benchmarks and kernels   so that they could either be evaluated on an  existing computer or the assist environment So let's talk about the testing evaluation process   this is a very important aspect to this program  that has been touched upon a couple times already The government test and evaluation team  will reproduce and thus validate significant   performer results validate all projections  experimental results designs and test suites   performers will submit their models and  designs for validation and performance using   assist achieving target metrics is one factor in  considering which teams move forward to the later   phase in the program proposals should be written  with the testing evaluation process in mind   testing protocols as written are preliminary and  will be finalized by the program kickoff meeting This chart provides information concerning  the task that was performed by the testing   evaluation teams they range from validating the  designs assessing the cops IP modules developing   the assist environment which as I said is based on  SST providing guidance on how to incorporate the   design into assist basically the performers are  developing a model within the assist environment   and evaluating the modifications to  the workflows kernels and benchmarks   I would not of course be surprised that a  some of the application software changes   the implementation of the algorithms based on  the change in the architecture so that seems   perfectly valid we just need to validate that it's  performing what is it supposed to be performing Okay so this is the program schedule for for phase  one as previously stated the first phase will last   18 months nine months after the start of  the program there will be a review of the   performers preliminary design at the 17th month  start of the 17th month performers will submit   their designs for test and evaluation during  this phase there will be several site visits   PI meetings and the obligatory monthly reports um  I imagine the way this program will be executed   though is there'll be constant interaction  between um the the design teams and myself   and the testing evaluation team so we really  won't be waiting until the 17th month to see what   the performance is I think we'll be marching  towards that as we go through this program Then for the phase two it's  very similar to the first phase   it's still 18 months long four months after  the start of this phase there will be a review   of the updated performers preliminary design  followed by a critical design review at 24 months   at the start of the 35th month performers will  submit their designs for tests and evaluation   and again numerous site visits and such I do want  to point out something I skipped over on the phase   one it's really important that these designs come  in at the beginning of the 17th month I will not   be allowed to advance to this next phase until the  results have been determined about the performance   of the workflows on the target with the workflows  and how they achieve the target metrics this is   an important consideration for the advancement of  the program so therefore delaying that will simply   delay the start of phase two for a particular  performer which is another reason why throughout   the program we want to have constant interaction  between the performers and the AGILE team So I'm almost done I talk too fast sorry about  that so you'll have some extra time to assemble   questions anyway so the chasm between the demands  of today's escalating data intensive problems and   the capabilities of of yesterday's computer system  is unbridgeable evolutionary improvements are now   providing diminishing routines AGILE is the first  program in decades to offer a clean sheet approach   for completely rethinking system level computing  architectures AGILE systems will enable a new era   of data analytics applications turning chaos into  order AGILE offers you an opportunity to pioneer   a new class of high-performance systems  that will change the face of computing   to be the next pioneer in computing who  changes the direction of computing I   think this is a very unique challenge and I am  hoping to receive some very exciting proposals Okay so thank you for listening to me describe  the AGILE program I encourage you to propose   a new exciting architecture that can genuinely  have an impact on computing carefully read the   baa write a proposal that fully responds to the  BAA when describing the proposed architecture   include a description how the design  impacts workflows and benchmarks provide   an estimate of the target metrics and finally  be realistic when considering the cost proposal   I have never seen a program at IARPA  or DARPA where the program manager   was able to get more funds for the program  because proposers didn't include enough   enough budgeted time into their program so please  do not do this if something isn't understood when   reading the BAA then please submit a question  to the AGILE BAA email address I believe that   this program provides a unique opportunity  seize the moment thank you for listening to me   okay sorry for the technical difficulties this  is not what my job discussion is about okay   good afternoon welcome back to the AGILE  Proposers' Day um our contracting officer   Chris Fox had uh had a personal emergency today  and she regrets not being able to be here to   actually go over this lives with you I am not a  contracting officer I do not play one on tv so I   am going to pretty much read this line if you  have questions you are welcome to ask them in   the questions section of this WebEx and we will  try to answer them and if we cannot answer them   then we will actually forward your questions  to a contracting officer who will be giving you   the official answer on behalf of the agency okay  with that I'm going to go to the next slide maybe   okay there we go so this is this is going to be  a Broad Agency Announcement um I'm going to try   to right now to to do questions and answers uh  regarding to um regarding to this possible BAA   who's eligible for applying how to prepare  the proposal how to submit the proposal how   the proposals will be evaluated and how the awards  process will take place okay we'll talk a little   bit about organization conflict of interest  intellectual property pre-publication reviews   academic institution acknowledgements  multiple proposal submissions and   contract types and there will be  of course a disclaimer at the end So IARPA uses the BAA-type solicitations  conducted under the far part 35 research   and development contracting okay um you know  where to go and look and look up the far and   on on the internet and you can  see what part 35 is all about   the BAA will be posted on beta.sam.gov which  is the place where all our baas are posted okay   we typically allow we typically allow between 45  and 60 days for proposals after the BAA has been   issued we are planning on this particular case  to do I believe 45 working days which turns out   to be about two months calendar uh by the time we  are all done with the period for proposal writing   all the information needed to submit a proposal  for this opportunity will be included in the BAA   nothing that I have said today or Bill has  said today supersedes the material in the BAA   so when in doubt go to the BAA the BAA will have a  question answered period during which prospective   offers can submit questions um that will be  very clearly stated when we announce the BAA   and you can submit your questions the emails  for the question the email to where to submit   the questions will be provided in the BAA the  questions will be provided in beta.sam.gov  

so be sure to check regularly for  the answers to your questions okay   no answers will go directly to the offer  nor shall questions be sent other than   than the email designator for the BAA and  know that your questions will be posted at   beta.sam.gov so if you think that you that  there might be some proprietary information   in your question please be very careful  because we will make all your questions public   and all answers to all your questions will also be  public okay so be careful about that collaboration   and teaming are generally encouraged by IARPA okay  so from all the people in this call if you guys   want to get together and make a single proposal  because you think that a team proposal will be   more solid please we encourage you to do that okay  however we do not facilitate for you to actually   uh for you to actually team that is up to you  to actually come up and team with other people   um foreign organizations and individuals  let's talk a little bit about foreign people   um this is the program dependent uh clauses and  the BAA the BAA will specify the limitation okay   there are other programs in path where we allow  participation from foreign nationals and foreign   entities okay regardless of eligibility any  foreign organization or individuals must comply   with any contract security clauses or requirements  and any united states export controls okay and   implementing contract contract clauses if you have  doubts about that please send your questions about   foreign foreign participation to us and we  will try to answer them once we open the BAA   the following are generally not eligible to  submit proposals for IARPA research programs   or participate as team members on the  proposals submitted by eligible entities   other government agencies we do not allow  government scientists to respond to our BAAs okay   FFRDC's we do not allow FFRDC entity employees  staff to actually be performers in our BAAs we do   not allow UARCs to be performers in our BAA okay  uh there's a clause there for you are an entity of   which only a portion has been designated at you  or may be eligible to submit subjects to an um   OCI review is started it is stated in the BAA okay  um so let's take a university that only a portion   of that university has been designated at New York  elements of that university outside of the New   York might actually be able to propose against the  BAA and we will do a review a conflict of interest   to actually ascertain whether or not you can  participate other organizations that have special   relationships with the government that will be  given the access to privilege or proprietary   information access to government equipment or  real property so you see this is not really for   government people to participate this is either  for industry or for academia to participate   so I'm sorry but that's those are the rules  and that's how that's how we conduct business   the BAA contains proposed proposal preparation  instructions please pay attention to the   instructions okay if we say five pages and you  decide to send 10 pages the last five pages will   be redacted and it's import if the important  information that you wanted us to know about   what's in those lives last five pages it will  end up in the trash can okay so please be careful   following the preparation instructions on the BAA  the BAA will include the due date and the time   okay page limitations and format information to  be addressed in the proposal for example technical   cost and administrative templates were required  for post proposal attachments for example the   cover sheet the conflict of interest forms  academic institution acknowledgement id and   data rights assertions cause breakdowns the BAA  also contains the evaluation factors for a war   including the technical evaluation criteria for  example technical approach relevance to the IARPA   work plan experience of the key personnel resource  resource realism etc. okay so you will know by   which metrics your proposal will be judged the PH  describes a method for evaluation and selection   the BAA may only request the technical volume  initially with the detailed cost volume requested   after selection so let me try to explain that  for those of you that are new we we might the BAA   might actually only be asking you for  the technical portion of your proposal   and if we like your technical idea and the  way in which you are trying to address it   we might then ask you to tell us  how much is it gonna cost okay   um but read the BAA the BAA will be very careful  about that so technical volume means the technical   portion of the proposal the cost volume  means the resource portion of your proposal   proposals must be submitted through IARPA  idea system the IARPA idea system if you   see the website there it's iarpa-ideas.gov  um you guys have to register electronically   into ideas before you can actually be able  to offer your proposal against a BAA okay   um right now if you're going to try to raise her  you will not be able to do it because the BAA is   not yet on the street so wait until the BAA hits  the street through beta.gov and then go into ideas  

register and then you will be  able to submit your proposal   insurance that offers are strongly encouraged to  register with ideas at least one week prior to   the proposal due date why in the past you know bad  things happen to computer systems in the past we   have had events in which ideas has gone down for  a couple of hours it just happens that it was the   last day of a BAA and you were desperate to try  to get in your proposal and now you're going to   register so take a week out of would we make maybe  registration at least a week before so if there's   some technical delicious we can actually help  you overcome them offers more most ensure that   that the version submitted to idea is the  final version um be careful double check your   document okay don't send us track changes  and stuff like that census sent us a clean   final version of your proposal if you decide  that your proposal is going to be classified   the BAA will explain to you how to send a  classified proposal to us okay um the BAA   will have instructions of how to respond in their  system pros with ideas I already said that before   just follow the instructions  that will take care of you okay   the cost volume is not requested until after  selection it will be directly submitted to the   contracting office officer not two ideas so  let me explain this again once again we are   looking for the technical portion of your proposal  okay that's the one that you send through ideas   a selection will be made based on the technical  merits of the proposals okay at that point you   are not at the other side of the rainbow yet  at that point you will enter negotiations   with the contracting officer and you will have to  submit the cost volume to the contracting officer   directly okay based on that negotiation with the  contracting officer you might end up having your   proposal rewarded with an award from IARPA okay  each BAA will detail the method for evaluation and   selection but IARPA generally follows a two-step  process first step is evaluation and selection   for negotiations this is conducted through a  scientific peer review process after which offers   are notified of selection I just said that like  a minute ago the second step is a negotiation and   contract award conducted by the contracting  officer I said that a minute ago okay   so we will deal with your technical idea first and  then we will deal with the contracting part okay   proposals will be reviewed individually against  the BAA requirements in accordance to 535 and not   against each other this is not a beauty contest we  take each one of your proposals on its own merits   if your proposal addresses  the requirements of the BAA   we will start a conversation  with you to follow up okay IARPA follows far part nine regarding  organizational conflicts of interest the main   principles being preventing conflict conflicting  rules that might bias a contractor's judgment and   preventing an unfair competitive advantage okay  the BAA will describe how offers are to identify   and disclose all disclose all facts related to  potential organizational conflicts of interest   for the offer as well as any proposed team members  it could be that in your team it's only one person   that might have an OCI then you disclose the  possible OCI of that person and we deal with   it one-on-one okay OCI disclosures may require  a mitigation plan describing the actions of the   offer that the offer will take or intends to  take to prevent the conflict in other words   just because you have an OCI doesn't mean  that we're going to kick you to the street   we will try to work with you to develop a  mitigation plan to try to take care of it so   you can still contribute to the to the BAA  IARPA generally prohibits contractors from   currently providing systems engineering technical  assistance theta and T&E support while being   technical R&D performers due to OCI concerns each  case will be determined in the beach individually   so like you remember from my presentation earlier  today we typically break the world between those   people that are proposing and performing against  our programs and those people that are doing   uh technical evaluation okay and we actually  separate the two so if you're playing on one   side of the fence you cannot play on the  other side of the fence and vice versa   there might be an incredibly weird case  in which we might allow you to do that but   we will determine that on an individual basis  the government needs to be able to effectively   manage the program and evaluate the output  and deliverables communicate the information   across government organizations and support  further use and development of program results   authors will address their intellectual property  rights assertions in their proposal the government   may request additional information as may be  necessary to evaluate those the government will   evaluate the IP rights being offered and whether  they are in the government's best interest in   other words protect your proper your intellectual  property okay we will help you do that but   your intellectual property is yours it's your  responsibility to help us help you protect it okay   IARPA encourages the publication of of  unclassified IARPA performer research   in peer-reviewed journals presentations at  conference and publication in conference   prior to the publication release of  any work submitted for publication   the performer will communicate the results to  be publicly released with the IARPA Program   Manager Bill Harrod in the scales to discuss any  sensitivity for example security speculation on   icus etc. and provide advanced currency copies to  the IARA program manager and contracting officer   of the publication there's a process  inside IARPA for the review of three pops   and I can tell you that until I sign off on the  pre-publications you will not be able to publish   that's the responsibility of office directors like  myself so we take this very seriously we address   your request on a timely manner and you are  able to public to to publish your your work okay   according to executive order 12333 contracts  or agreements with academic institutions may   be undertaken only with the consent of  appropriate officials of the institution   an academic institution acknowledgement letter is  required for offers that are academic institutions   and for any proposed teammate that is an academic  institution a template of this letter will be   included in the BAA each letter must be signed by  a senior official of the institution for example   the president the chancellor the provost  or other appropriate designated individual   IARPA requires this letter before entering  into negotiations and or awarding contract   it is highly advised that it be submitted with  the proposal ask permission to submit the proposal   from an official at the university before you  send it to us without that we cannot consider it   proposal submissions to other entities typically  the behavior the BAA has offers to name   in their proposal other federal state or local  agencies and other parties receiving the proposal   or substantially the same proposal for funding of  the proposed effort if the offer has submitted the   same or substantially the same proposal to other  entities it may impact decision to select and fund   the effort I promise you we talk to each other  if you send your proposal already to DARPA and   you're sending it to us again we will know that  okay so please avoid doing that multiple proposal   submissions to IARPA the BAA usually allow an  entity to participate in multiple submissions   as a prime or a subcontractor if allowed by the  BAA multiple submissions which include a common   team member shall not receive duplicate funding  for the same work I no no one entry can be paid   twice for the same work so if you want to go and  be the guy that develops gadget a for proposal one   and you want to be the guy that developed gadget  a for proposal two that is allowed however if   proposal one and proposal two are selected  you will not get paid twice for gadget a Cost or cost plus fixed fee type contracts are  typically awarded due to the nature of the R&D   work IARPA may in some instances consider  other contact types such as firm fixed price   as well as non-part-based agreements such as other  transactions the types of contracts and agreements   that will be considered and the conditions for  such consideration for example small business   startups commercial foreign entities etc.  will be addressed in the BAA so please read   the BAA careful about that the information  conveying this briefing is for planning   and general information purposes and is subject  to change please carefully read the final BAA   and adhere to its requirements which may differ  from what has been presented in this briefing in   other words my word or the word word of Bill today  is not the final word the final word is the BAA   and with that it's time for me to be quiet and  allow you guys to continue to send your questions   in there's an army of of IARPA uh staff  members right now reading your questions   answering them and then Bill will

2021-01-09

Show video