Causal AI for Systems Learning Causal Performance Models for conducting Performance Tasks in a Principled and Transferable Fashion Pooyan Jamshidi

What is Causal AI?

Let’s start with a (fiction) story • Zeus is a patient waiting for a heart transplant. On 1 January, he received a new heart. Five days later, he died. • Imagine that we can somehow know, that had Zeus not received a heart transplant on 1 January then he would have been alive fi ve days later. • All others things in his life being unchanged. • Now, what do you think was the cause of Zeus’s death?! • Most people would agree that the transplant caused Zeus’ death. • The intervention had a causal e ff ect.

Let’s start with a (fiction) story • Hera, received a heart transplant on 1 January. Five days later she was alive. • Again, imagine we can somehow know that had Hera not received the heart on 1 January then she would still have been alive fi ve days later. • All others things in his life being unchanged. • The transplant did not have a causal e ff ect on Hera’s fi ve day survival.

Let’s collect some data! Exposure variable A (1: exposed, 0: unexposed); Outcome variable Y (1: death, 0: survival)

Individual Causal Effect contrast of the values of counterfactual outcomes, but only one of those values is observed.

Population Causal Effects • Pr[Ya = 1]: proportion of subjects that would have developed the outcome Y had all subjects in the population of interest received exposure value a. • The exposure has a causal e ff ect in the population if 
 Pr[Ya=1=1] Pr[Ya=0=1]. • Unlike individual causal e ff ects, population causal e ff ects can sometimes be computed—or, more rigorously, consistently estimated. ≠ Pr[Ya=1 = 1] − Pr[Ya=0 = 1] ≠ 0

Now let’s do some cool ML ML models characterize association Pr[Y = 1|A = 1] = 7/13 Pr[Y = 1|A = 0] = 3/7

Association is not Causation!

Computing Causal Effects via Randomization Unlike association measures, e ff ect measures cannot be directly computed because of missing data. However, e ff ect measures can be computed/estimated in randomized experiments! • Suppose we have a (near-in fi nite) population and that we fl ip a coin for each subject in such population. We assign the subject to group 1 if the coin turns tails, and to group 2 if it turns heads. • Next we administer the treatment or exposure of interest (A = 1) to subjects in group 1 and placebo (A = 0) to those in group 2. Five days later, at the end of the study, we compute the mortality risks in each group, Pr[Y = 1|A = 1] and Pr[Y = 1|A = 0]. • When subjects are randomly assigned to groups 1 and 2, the proportion of deaths among the exposed, Pr[Y = 1|A = 1], will be the same whether subjects in group 1 receive the exposure and subjects in group 2 receive placebo, or vice versa. • Because group membership is randomised, both groups are ‘‘comparable’’: which particular group got the exposure is irrelevant for the value of Pr[Y = 1|A = 1]. (The same reasoning applies to Pr[Y = 1|A = 0].) • Formally, we say that both groups are exchangeable.

Let’s do some math! Pr[Y = 1|A = 1] = Pr[Y = 1|A = 0] = Pr[Ya = 1] Pr[Ya = 1|A = a] = Pr[Y = 1|A = a] Pr[Y = 1|A = a] = Pr[Ya = 1] In ideal randomized experiments, Association is Causation!

But not in non-randomized observational studies Still remember this? Pr[Y = 1|A = 1] = 7/13 Pr[Y = 1|A = 0] = 3/7

Outline 13 Causal AI For Systems UNICORN Results Future Directions Motivation

14 Goal: Enable developers/users to fi nd the right quality tradeoff

Today’s most popular systems are con fi gurable 15 built

16

Empirical observations con fi rm that systems are becoming increasingly con fi gurable 17 08 7/2010 7/2012 7/2014 Release time 1/1999 1/2003 1/2007 1/2011 0 1/2014 N Release time 02 1/2006 1/2010 1/2014 2.2.14 2.3.4 2.0.35 .3.24 Release time Apache 1/2006 1/2008 1/2010 1/2012 1/2014 0 40 80 120 160 200 2.0.0 1.0.0 0.19.0 0.1.0 Hadoop Number of parameters Release time MapReduce HDFS [Tianyin Xu, et al., “Too Many Knobs…”, FSE’15]

Empirical observations con fi rm that systems are becoming increasingly con fi gurable 18 nia San Diego, ‡Huazhong Univ. of Science & Technology, †NetApp, Inc tixu, longjin, xuf001, yyzhou}@cs.ucsd.edu kar.Pasupathy, Rukma.Talwadker}@netapp.com prevalent, but also severely software. One fundamental y of configuration, reflected parameters (“knobs”). With m software to ensure high re- aunting, error-prone task. nderstanding a fundamental users really need so many answer, we study the con- including thousands of cus- m (Storage-A), and hundreds ce system software projects. ng findings to motivate soft- ore cautious and disciplined these findings, we provide ich can significantly reduce A as an example, the guide- ters and simplify 19.7% of on existing users. Also, we tion methods in the context 7/2006 7/2008 7/2010 7/2012 7/2014 0 100 200 300 400 500 600 700 Storage-A Number of parameters Release time 1/1999 1/2003 1/2007 1/2011 0 100 200 300 400 500 5.6.2 5.5.0 5.0.16 5.1.3 4.1.0 4.0.12 3.23.0 1/2014 MySQL Number of parameters Release time 1/1998 1/2002 1/2006 1/2010 1/2014 0 100 200 300 400 500 600 1.3.14 2.2.14 2.3.4 2.0.35 1.3.24 Number of parameters Release time Apache 1/2006 1/2008 1/2010 1/2012 1/2014 0 40 80 120 160 200 2.0.0 1.0.0 0.19.0 0.1.0 Hadoop Number of parameters Release time MapReduce HDFS [Tianyin Xu, et al., “Too Many Knobs…”, FSE’15]

Con fi gurations determine the performance behavior 19 void Parrot_setenv(. . . name,. . . value){ #ifdef PARROT_HAS_SETENV my_setenv(name, value, 1); #else int name_len=strlen(name); int val_len=strlen(value); char* envs=glob_env; if(envs==NULL){ return; } strcpy(envs,name); strcpy(envs+name_len,"="); strcpy(envs+name_len + 1,value); putenv(envs); #endif } #ifdef LINUX extern int Parrot_signbit(double x){ endif else PARROT_HAS_SETENV LINUX Speed Energy

Misconfiguration and its Effects ● Misconfigurations can elicit unexpected interactions between software and hardware ● These can result in non-functional faults ○ Affecting non-functional system properties like latency, throughput, energy consumption, etc. 20 The system doesn’t crash or exhibit an obvious misbehavior Systems are still operational but with a degraded performance, e.g., high latency, low throughput, high energy consumption, high heat dissipation, or a combination of several

21 CUDA performance issue on tx2 When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. The user is transferring the code from one hardware to another When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. The target hardware is faster than the the source hardware. User expects the code to run at least 30-40% faster. Motivating Example When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. The code ran 2x slower on the more powerful hardware

Motivating Example 22 June 3rd We have already tried this. We still have high latency. Any other suggestions? June 4th Please do the following and let us know if it works 1. Install JetPack 3.0 2. Set nvpmodel=MAX-N 3. Run jetson_clock.sh June 5th June 4th TX2 is pascal architecture. Please update your CMakeLists: + set(CUDA_STATIC_RUNTIME OFF) ... + -gencode=arch=compute_62,code=sm_62 The user had several misconfigurations In Software: ✖ Wrong compilation flags ✖ Wrong SDK version In Hardware: ✖ Wrong power mode ✖ Wrong clock/fan settings The discussions took 2 days ! Any suggestions on how to improve my performance? Thanks! How to resolve such issues faster? ?

Today’s most popular systems are complex! multiscale, multi-modal, and multi-stream 23 Variability Space = Con fi guration Space + System Architecture + Deployment Environment Video Decoder Stream Muxer Primary Detector Object Tracker Secondary Classifier # Configuration Options 55 86 14 44 86

24 More combinations than estimated atoms in the universe

0 500 1000 1500 Throughput (ops/sec) 0 1000 2000 3000 4000 5000 Average write latency ( s) The default con fi guration is typically bad and the optimal con fi guration is noticeably better than median 25 Default Con fi guration Optimal Con fi guration better better • Default is bad • 2X-10X faster than worst • Noticeably faster than median

Performance behavior varies in different environments 26

Outline 27 Motivation UNICORN Results Future Directions Causal AI For Systems

Causal AI in Systems and Software 28 Computer Architecture Database Operating Systems Programming Languages BigData Software Engineering https://github.com/y-ding/causal-system-papers

29 Throughput = 9 × Bitrate + 2.1 × Buffersize − 4.4 × Bitrate × Buffersize × BatchSize Causal Performance Model Traditional Performance Model VS Throughput Energy Branch Misses Cache Misses No. of Cycles Bitrate Buffer Size Batch Size Enable Padding f3 f4 f f1 f2 Causal Interaction Causal Paths Software Options Intermediate Causal Mechanisms Performance Objective f Branchmisses = 2 × Bitrate + 8.1 × Buffersize + 4.1 × Bitrate × Buffersize × Cachemisses Decoder Muxer

Critical Issues of Correlation-based Performance Analysis • Performance in fl uence models could produce unreliable predictions. • Performance in fl uence models could produce unstable predictions across environments and in the presence of measurement noise. • Performance in fl uence models could produce incorrect explanations. 30

Why Causal Inference? (Simpson’s Paradox) 31 Increasing GPU memory increases Latency More GPU memory usage should reduce latency not increase it. Counterintuitive! Any ML-/statistical models built on this data will be incorrect !

Why Causal Inference? (Simpson’s Paradox) 32 Segregate data on swap memory Available swap memory is reducing GPU memory borrows memory from the swap for some intensive workloads. Other host processes may reduce the available swap. Little will be left for the GPU to use.

33 Why Causal Inference? Real world problems can have 100s if not 1000s of interacting configuration options ! Manually understanding and evaluating each combination is impractical, if not impossible.

Load GPU Mem. Swap Mem. Latency Express the relationships between interacting variables as a causal graph 34 Causal Performance Models Configuration option Direction(s) of the causality • Latency is affected by GPU Mem. which in turn is influenced by swap memory • External factors like resource pressure also affects swap memory Non-functional property System event

35 Causal Performance Models How to construct this causal graph? ? If there is a fault in latency, how to diagnose and fix it? ? Load GPU Mem. Swap Mem. Latency

No content

Outline 37 Motivation Causal AI For Systems Results Future Directions UNICORN

• Build a Causal Performance Model that capture the interactions options in the variability space using the observation performance data. • Iterative causal performance model evaluation and model update • Perform downstream performance tasks such as performance debugging & optimization using Causal Reasoning UNICORN: Our Causal AI for Systems Method

UNICORN: Our Causal AI for Systems Method Software: DeepStream Middleware: TF, TensorRT Hardware: Nvidia Xavier Configuration: Default number of counters number of splitters latency (ms) 100 150 1 200 250 2 300 Cubic Interpolation Over Finer Grid 2 4 3 6 8 4 10 12 5 14 16 6 18 Budget Exhausted? Yes No 5- Update Causal Performance Model Query Engine 4- Estimate Causal Queries Estimate probability of satisfying QoS if BufferSize is set to 6k? 2- Learn Causal Performance Model Performance Debugging Performance Optimization 3- Translate Perf. Query to Causal Queries •What is the root-cause of observed perf. fault? •How do I fix the misconfig.? •How can I improve throughput without sacrificing accuracy? •How do I understand perf behavior? Measure performance of the configuration(s) that maximizes information gain Performance Data Causal Model P(Th > 40/s|do(Buffersize = 6k)) 1- Specify Performance Query QoS : Th > 40/s Observed : Th < 30/s ± 5/s

Software: DeepStream Middleware: TF, TensorRT Hardware: Nvidia Xavier Configuration: Default number of counters number of splitters latency (ms) 100 150 1 200 250 2 300 Cubic Interpolation Over Finer Grid 2 4 3 6 8 4 10 12 5 14 16 6 18 Budget Exhausted? Yes No 5- Update Causal Performance Model Query Engine 4- Estimate Causal Queries Estimate probability of satisfying QoS if BufferSize is set to 6k? 2- Learn Causal Perf. Model Performance Debugging Performance Optimization 3- Translate Performance Query to Causal Queries •What is the root-cause of observed perf. fault? •How do I fix the misconfig.? •How can I improve throughput without sacrificing accuracy? •How do I understand perf behavior? Measure performance of the configuration(s) that maximizes information gain Performance Data Causal Model P(Th > 40/s|do(Buffersize = 6k)) 1- Specify Performance Query QoS : Th > 40/s Observed : Th < 30/s ± 5/s UNICORN: Our Causal AI for Systems Method

Software: DeepStream Middleware: TF, TensorRT Hardware: Nvidia Xavier Configuration: Default number of counters number of splitters latency (ms) 100 150 1 200 250 2 300 Cubic Interpolation Over Finer Grid 2 4 3 6 8 4 10 12 5 14 16 6 18 Budget Exhausted? Yes No 5- Update Causal Performance Model Query Engine 4- Estimate Causal Queries Estimate probability of satisfying QoS if BufferSize is set to 6k? 2- Learn Causal Perf. Model Performance Debugging Performance Optimization 3- Translate Performance Query to Causal Queries •What is the root-cause of observed perf. fault? •How do I fix the misconfig.? •How can I improve throughput without sacrificing accuracy? •How do I understand perf behavior? Measure performance of the configuration(s) that maximizes information gain Performance Data Causal Model P(Th > 40/s|do(Buffersize = 6k)) 1- Specify Performance Query QoS : Th > 40/s Observed : Th < 30/s ± 5/s UNICORN: Our Causal AI for Systems Method

FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding Bitrate (bits/s) Enable Padding … Cache Misses … Through put (fps) c1 1k 1 … 42m … 7 c2 2k 1 … 32m … 22 … … … … … … … cn 5k 0 … 12m … 25 FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding 1- Recovering the Skelton 2- Pruning Causal Structure 3- Orienting Causal Relations statistical independence tests fully connected graph given constraints (e.g., no connections btw configuration options) orientation rules & measures (entropy) + structural constraints (colliders, v-structures) Learning Causal Performance Model

Performance measurement 43 ℂ = O1 × O2 × ⋯ × O19 × O20 Dead code removal Con fi guration Space Constant folding Loop unrolling Function inlining c1 = 0 × 0 × ⋯ × 0 × 1 c1 ∈ ℂ fc (c1 ) = 11.1ms Compile time Execution time Energy Compiler (e.f., SaC, LLVM) Program Compiled Code Instrumented Binary Hardware Compile Deploy Con fi gure fe (c1 ) = 110.3ms fen (c1 ) = 100mwh Non-functional measurable/quanti fi able aspect

Our setup for performance measurements 44

Hardware platforms in our experiments The reason behind using di ff erent types of hardware platforms is that they exhibit di ff erent behaviors due to di ff erences in terms of resources, their microarchitecture, etc. 45 AWS DeepLens: Cloud-connected device System on Chip (SoC) Microcontrollers (MCUs)

Measuring performance for systems involves lots of challenges Each hardware requires di ff erent ways of instrumentations and clean measurement that contains least amount of noise is the most challenging part of our experiments. 46

FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding Bitrate (bits/s) Enable Padding … Cache Misses … Through put (fps) c1 1k 1 … 42m … 7 c2 2k 1 … 32m … 22 … … … … … … … cn 5k 0 … 12m … 25 FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding 1- Recovering the Skelton 2- Pruning Causal Structure 3- Orienting Causal Relations statistical independence tests fully connected graph given constraints (e.g., no connections btw configuration options) orientation rules & measures (entropy) + structural constraints (colliders, v-structures) Learning Causal Performance Model

FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding Bitrate (bits/s) Enable Padding … Cache Misses … Through put (fps) c1 1k 1 … 42m … 7 c2 2k 1 … 32m … 22 … … … … … … … cn 5k 0 … 12m … 25 FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding 1- Recovering the Skelton 2- Pruning Causal Structure 3- Orienting Causal Relations statistical independence tests fully connected graph given constraints (e.g., no connections btw configuration options) orientation rules & measures (entropy) + structural constraints (colliders, v-structures) Learning Causal Performance Model

FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding Bitrate (bits/s) Enable Padding … Cache Misses … Through put (fps) c1 1k 1 … 42m … 7 c2 2k 1 … 32m … 22 … … … … … … … cn 5k 0 … 12m … 25 FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding 1- Recovering the Skelton 2- Pruning Causal Structure 3- Orienting Causal Relations statistical independence tests fully connected graph given constraints (e.g., no connections btw configuration options) orientation rules & measures (entropy) + structural constraints (colliders, v-structures) Learning Causal Performance Model

Throughput Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding f f f f f Causal Interaction Causal Paths Software Options Perf. Events Performance Objective f Branchmisses = 2 × Bitrate + 8.1 × Buffersize + 4.1 × Bitrate × Buffersize × Cachemisses Decoder Muxer Causal Performance Model

Software: DeepStream Middleware: TF, TensorRT Hardware: Nvidia Xavier Configuration: Default number of counters number of splitters latency (ms) 100 150 1 200 250 2 300 Cubic Interpolation Over Finer Grid 2 4 3 6 8 4 10 12 5 14 16 6 18 Budget Exhausted? Yes No 5- Update Causal Performance Model Query Engine 4- Estimate Causal Queries Estimate probability of satisfying QoS if BufferSize is set to 6k? 2- Learn Causal Perf. Model Performance Debugging Performance Optimization 3- Translate Performance Query to Causal Queries •What is the root-cause of observed perf. fault? •How do I fix the misconfig.? •How can I improve throughput without sacrificing accuracy? •How do I understand perf behavior? Measure performance of the configuration(s) that maximizes information gain Performance Data Causal Model P(Th > 40/s|do(Buffersize = 6k)) 1- Specify Performance Query QoS : Th > 40/s Observed : Th < 30/s ± 5/s UNICORN: Our Causal AI for Systems Method

52 Diagnose and fix the root-cause of misconfigurations that cause non-functional faults Objective Causal Debugging: An example of downstream performance task Ὂ Use causal models to model various cross-stack configuration interactions; and Ὂ Counterfactual reasoning to recommend fixes for these misconfigurations Approach

53 Causal Debugging • What is the root-cause of my fault? • How do I fix my misconfigurations to improve performance? Misconfiguration Fault fixed? Observational Data Build Causal Graph Extract Causal Paths Best Query Yes No update observational data Counterfactual Queries Rank Paths What if questions. E.g., What if the configuration option X was set to a value ‘x’? About 25 sample configurations (training data)

Best Query Counterfactual Queries Rank Paths What if questions. E.g., What if the configuration option X was set to a value ‘x’? Extract Causal Paths 54 Extracting Causal Paths from the Causal Model • What is the root-cause of my fault? • How do I fix my misconfigurations to improve performance? Misconfiguration Fault fixed? Observational Data Build Causal Graph Yes No update observational data About 25 sample configurations (training data)

Extracting Causal Paths from the Causal Model Problem ✕ In real world cases, this causal graph can be very complex ✕ It may be intractable to reason over the entire graph directly 55 Solution ✓ Extract paths from the causal graph ✓ Rank them based on their Average Causal Effect on latency, etc. ✓ Reason over the top K paths

Extracting Causal Paths from the Causal Model 56 GPU Mem. Latency Swap Mem. Extract paths Always begins with a configuration option Or a system event Always terminates at a performance objective Load GPU Mem. Latency Swap Mem. Swap Mem. Latency Load GPU Mem.

Ranking Causal Paths from the Causal Model 57 ● They may be too many causal paths ● We need to select the most useful ones ● Compute the Average Causal Effect (ACE) of each pair of neighbors in a path GPU Mem. Swap Mem. Latency 𝐴𝐶 𝐸 (GPU Mem . , Swap) = 1 𝑁 ∑ 𝑎 , 𝑏 ∈ 𝑍 𝔼 (GPU Mem . 𝑑 𝑜 (Swap = 𝑏 )) − 𝔼 (GPU Mem . 𝑑 𝑜 (Swap = 𝑎 )) Expected value of GPU Mem. when we artificially intervene by setting Swap to the value b Expected value of GPU Mem. when we artificially intervene by setting Swap to the value a If this difference is large, then small changes to Swap Mem. will cause large changes to GPU Mem. Average over all permitted values of Swap memory.

Ranking Causal Paths from the Causal Model 58 ● Average the ACE of all pairs of adjacent nodes in the path ● Rank paths from highest path ACE (PACE) score to the lowest ● Use the top K paths for subsequent analysis 𝑃𝐴𝐶𝐸 ( 𝑍 , 𝑌 ) = 1 2 ( 𝐴 𝐶 𝐸 ( 𝑍 , 𝑋 ) + 𝐴𝐶 𝐸 ( 𝑋 , 𝑌 )) X Y Z Sum over all pairs of nodes in the causal path. GPU Mem. Latency Swap Mem.

Best Query Counterfactual Queries Rank Paths What if questions. E.g., What if the configuration option X was set to a value ‘x’? Extract Causal Paths 59 Diagnosing and Fixing the Faults • What is the root-cause of my fault? • How do I fix my misconfigurations to improve performance? Misconfiguration Fault fixed? Observational Data Build Causal Graph Yes No update observational data About 25 sample configurations (training data)

Diagnosing and Fixing the Faults 60 ● Counterfactual inference asks “what if” questions about changes to the misconfigurations We are interested in the scenario where: • We hypothetically have low latency; Conditioned on the following events: • We hypothetically set the new Swap memory to 4 Gb • Swap Memory was initially set to 2 Gb • We observed high latency when Swap was set to 2 Gb • Everything else remains the same Example Given that my current swap memory is 2 Gb, and I have high latency. What is the probability of having low latency if swap memory was increased to 4 Gb?

Low? Load GPU Mem. Latency Swap = 4 Gb Diagnosing and Fixing the Faults 61 GPU Mem. Latency Swap Original Path Load GPU Mem. Latency Swap = 4 Gb Path after proposed change Load Remove incoming edges. Assume no external influence. Modify to reflect the hypothetical scenario Low? Load GPU Mem. Latency Swap = 4 Gb Low? Use both the models to compute the answer to the counterfactual question

Diagnosing and Fixing the Faults 62 GPU Mem. Latency Swap Original Path Load GPU Mem. Latency Swap = 4 Gb Path after proposed change Load 𝑃 𝑜 𝑡 𝑒 𝑛 𝑡𝑖 𝑎 𝑙 = 𝑃 ( ^ 𝐿𝑎 𝑡 𝑒 𝑛𝑐 𝑦 = 𝑙 𝑜𝑤 . . ^ 𝑆𝑤 𝑎𝑝 = 4 𝐺 𝑏 , . 𝑆 𝑤 𝑎𝑝 = 2 𝐺 𝑏 , 𝐿𝑎 𝑡 𝑒 𝑛𝑐𝑦 𝑠 𝑤 𝑎 𝑝 =2 𝐺 𝑏 = h 𝑖𝑔 h, 𝑈 ) We expect a low latency The latency was high The Swap is now 4 Gb The Swap was initially 2 Gb Everything else stays the same

Diagnosing and Fixing the Faults 63 Potential = 𝑃 ( ^ 𝑜𝑢𝑡𝑐𝑜𝑚 𝑒 = 𝑔𝑜 𝑜𝑑 ~ ~ 𝑐 h 𝑎 𝑛 𝑔 𝑒 , ~ 𝑜 𝑢 𝑡𝑐𝑜 𝑚 𝑒 ¬ 𝑐 h 𝑎 𝑛 𝑔 𝑒 = 𝑏𝑎𝑑 , ~¬ 𝑐 h 𝑎 𝑛 𝑔𝑒 , 𝑈 ) Probability that the outcome is good after a change, conditioned on the past If this difference is large, then our change is useful Individual Treatment Effect = Potential − Outcome Control = 𝑃 ( ^ 𝑜𝑢 𝑡 𝑐 𝑜 𝑚 𝑒 = 𝑏𝑎𝑑 ~ ~¬ 𝑐 h 𝑎 𝑛𝑔 𝑒 , 𝑈 ) Probability that the outcome was bad before the change

Diagnosing and Fixing the Faults 64 GPU Mem. Latency Swap Mem. Top K paths ⋮ Enumerate all possible changes 𝐼 𝑇 𝐸 ( 𝑐 h 𝑎𝑛𝑔 𝑒 ) Change with the largest ITE Set every configuration option in the path to all permitted values Inferred from observed data. This is very cheap. !

Diagnosing and Fixing the Faults 65 Change with the largest ITE Fault fixed? Yes No • Add to observational data • Update causal model • Repeat… Measure Performance

Software: DeepStream Middleware: TF, TensorRT Hardware: Nvidia Xavier Configuration: Default number of counters number of splitters latency (ms) 100 150 1 200 250 2 300 Cubic Interpolation Over Finer Grid 2 4 3 6 8 4 10 12 5 14 16 6 18 Budget Exhausted? Yes No 5- Update Causal Performance Model Query Engine 4- Estimate Causal Queries Estimate probability of satisfying QoS if BufferSize is set to 6k? 2- Learn Causal Perf. Model Performance Debugging Performance Optimization 3- Translate Performance Query to Causal Queries •What is the root-cause of observed perf. fault? •How do I fix the misconfig.? •How can I improve throughput without sacrificing accuracy? •How do I understand perf behavior? Measure performance of the configuration(s) that maximizes information gain Performance Data Causal Model P(Th > 40/s|do(Buffersize = 6k)) 1- Specify Performance Query QoS : Th > 40/s Observed : Th < 30/s ± 5/s UNICORN: Our Causal AI for Systems Method

FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding 1- Evaluate Candidate Interventions FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding Option/Event/Obj Values Bitrate 1k Buffer Size 20k Batch Size 10 Enable Padding 1 Branch Misses 24m Cache Misses 42m No of Cycles 73b FPS 31/s Energy 42J 2- Determine & Perform next Perf Measurement 3- Updating Causal Model Performance Data Model averaging Expected change in belief & KL; Causal effects on objectives Interventions on Hardware, Workload, and Kernel Options Active Learning for Updating Causal Performance Model

FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding 1- Evaluate Candidate Interventions FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding Option/Event/Obj Values Bitrate 1k Buffer Size 20k Batch Size 10 Enable Padding 1 Branch Misses 24m Cache Misses 42m No of Cycles 73b FPS 31/s Energy 42J 2- Determine & Perform next Perf Measurement 3- Updating Causal Model Performance Data Model averaging Expected change in belief & KL; Causal effects on objectives Interventions on Hardware, Workload, and Kernel Options Active Learning for Updating Causal Performance Model

FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding 1- Evaluate Candidate Interventions FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enable Padding Option/Event/Obj Values Bitrate 1k Buffer Size 20k Batch Size 10 Enable Padding 1 Branch Misses 24m Cache Misses 42m No of Cycles 73b FPS 31/s Energy 42J 2- Determine & Perform next Perf Measurement 3- Updating Causal Model Performance Data Model averaging Expected change in belief & KL; Causal effects on objectives Interventions on Hardware, Workload, and Kernel Options Active Learning for Updating Causal Performance Model

Benefits of Causal Reasoning for System Performance Analysis

There are two fundamental benefits that we get by our “Causal AI for Systems” methodology 1. We learn one central (causal) performance model from the data across di ff erent performance tasks: • Performance understanding • Performance optimization • Performance debugging and repair • Performance prediction for di ff erent environments (e.g., canary-> production) 2. The causal model is transferable across environments. • We observed Sparse Mechanism Shift in systems too! • Alternative non-causal models (e.g., regression-based models for performance tasks) are not transferable as they rely on i.i.d. setting. 71

Questions of this nature require precise mathematical language lest they will be misleading. Here we are simultaneously conditioning on two values of GPU memory growth (i.e., 𝑋 ˆ = 0.66 and 𝑋 = 0.33). Traditional machine learning approaches cannot handle such expressions. Instead, we must resort to causal models to compute them. 72

Difference between statistical (left) and causal models (right) on a given set of three variables While a statistical model speci fi es a single probability distribution, a causal model represents a set of distributions, one for each possible intervention. 73

Independent Causal Mechanisms (ICM) Principle

Sparse Mechanism Shift (SMS) Hypothesis Example of SMS hypothesis, where an intervention (which may or may not be intentional/observed) changes the position of one fi nger, and as a consequence, the object falls. The change in pixel space is entangled (or distributed), in contrast to the change in the causal model.

76 NeurIPS 2020 (ML For Systems), Dec 12th, 2020 https://arxiv.org/pdf/2010.06061.pdf https://github.com/softsys4ai/CADET

77 The new version of CADET, called UNICORN, accepted at EuroSys 2022. https://github.com/softsys4ai/UNICORN

Outline 78 Motivation Causal AI For Systems Future Directions UNICORN Results

Results: Case Study 79 When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. When we are trying to transplant our CUDA source code from TX1 to TX2, it behaved strange. We noticed that TX2 has twice computing-ability as TX1 in GPU, as expectation, we think TX2 will 30% - 40% faster than TX1 at least. Unfortunately, most of our code base spent twice the time as TX1, in other words, TX2 only has 1/2 speed as TX1, mostly. We believe that TX2’s CUDA API runs much slower than TX1 in many cases. The user is transferring the code from one hardware to another The target hardware is faster than the the source hardware. User expects the code to run at least 30-40% faster. The code ran 2x slower on the more powerful hardware

More powerful Results: Case Study 80 Nvidia TX1 CPU 4 cores, 1.3 GHz GPU 128 Cores, 0.9 GHz Memory 4 Gb, 25 Gb/s Nvidia TX2 CPU 6 cores, 2 GHz GPU 256 Cores, 1.3 GHz Memory 8 Gb, 58 Gb/s Embedded real-time stereo estimation Source code 17 Fps 4 Fps 4 Slower! ×

Results: Case Study 81 Configuration UNICO RN Decision Tree Forum CPU Cores ✓ ✓ ✓ CPU Freq. ✓ ✓ ✓ EMC Freq. ✓ ✓ ✓ GPU Freq. ✓ ✓ ✓ Sched. Policy ✓ Sched. Runtime ✓ Sched. Child Proc ✓ Dirty Bg. Ratio ✓ Drop Caches ✓ CUDA_STATIC_RT ✓ ✓ ✓ Swap Memory ✓ UNICORN Decision Tree Forum Throughput (on TX2) 26 FPS 20 FPS 23 FPS Throughput Gain (over TX1) 53 % 21 % 39 % Time to resolve 24 min. 31/2 Hrs. 2 days X Finds the root-causes accurately X No unnecessary changes X Better improvements than forum’s recommendation X Much faster Results The user expected 30-40% gain

Evaluation: Experimental Setup Nvidia TX1 CPU 4 cores, 1.3 GHz GPU 128 Cores, 0.9 GHz Memory 4 Gb, 25 GB/s Nvidia TX2 CPU 6 cores, 2 GHz GPU 256 Cores, 1.3 GHz Memory 8 Gb, 58 GB/s Nvidia Xavier CPU 8 cores, 2.26 GHz GPU 512 cores, 1.3 GHz Memory 32 Gb, 137 GB/s Hardware Systems Software Systems Xception Image recognition (50,000 test images) DeepSpeech Voice recognition (5 sec. audio clip) BERT Sentiment Analysis (10000 IMDb reviews) x264 Video Encoder (11 Mb, 1080p video) Configuration Space X 30 Configurations X 17 System Events • 10 software • 10 OS/Kernel • 10 hardware 82

Evaluation: Data Collection ● For each software/hardware combination create a benchmark dataset ○ Exhaustively set each of configuration option to all permitted values. ○ For continuous options (e.g., GPU memory Mem.), sample 10 equally spaced values between [min, max] ● Measure the latency, energy consumption, and heat dissipation ○ Repeat 5x and average 83 Multiple Faults ! Latency Faults ! Energy Faults !

Evaluation: Ground Truth ● For each performance fault: ○ Manually investigate the root-cause ○ “Fix” the misconfigurations ● A “fix” implies the configuration no longer has tail performance ○ User defined benchmark (i.e., 10th percentile) ○ Or some QoS/SLA benchmark ● Record the configurations that were changed 84 Multiple Faults ! Latency Faults ! Energy Faults !

Evaluation: Metrics 85 Relevance Scores 𝐺 𝑎 𝑖 𝑛 = NFP fault − NFP repair NFP fault × 100 Repair Quality NFP = Non-Functional Property (e.g., Latency, Energy, etc.) Repair value Faulty value Larger the gain, better the repair

RQ2: How does UNICORN perform compared to Search-Based Optimization 86 RQ1: How does UNICORN perform compared to Model based Diagnostics Results: Research Questions

87 Results: Research Question 1 (single objective) RQ1: How does UNICORN perform compared to Model based Diagnostics X Finds the root-causes accurately X Better gain X Much faster Takeaways More accurate than ML-based methods Better Gain Up to 20x faster

88 Results: Research Question 1 (multi-objective) RQ1: How does UNICORN perform compared to Model based Diagnostics X No deterioration of other performance objectives Takeaways Multiple Faults in Latency & Energy usage

RQ1: How does UNICORN perform compared to Model based Diagnostics 89 RQ2: How does UNICORN perform compared to Search-Based Optimization Results: Research Questions

Results: Research Question 2 RQ2: How does UNICORN perform compared to Search-Based Optimization X Better with no deterioration of other performance objectives Takeaways 90

91 Results: Research Question 3 RQ2: How does UNICORN perform compared to Search-Based Optimization X Considerably faster than search-based optimization Takeaways

Outline 92 Motivation Causal AI For Systems UNICORN Results Future Directions

Causal AI for Serverless • Evaluating our Causal AI for Systems methodology with Serverless systems provide the following opportunities: 1. Dynamic system recon fi gurations • Dynamic placement of functions • Dynamic recon fi gurations of the network of functions • Dynamic multi-cloud placement of functions. 2. Root cause analysis of failures or QoS drop 93

Causal AI for Autonomous Robot Testing • Testing cyberphysical systems such as robots are di ff i cult. The key reason is that there are additional interactions with the environment and the task that the robot is performing. • Evaluating our Causal AI for Systems methodology with autonomous robots provide the following opportunities: 1. Identifying di ff i cult to catch bugs in robots 2. Identifying the root cause of an observed fault and repairing the issue automatically during mission time. 94

Summary: Causal AI for Systems 1. Learning a Functional Causal Model for di ff erent downstream systems tasks 2. The learned causal model is transferable across di ff erent environments 95 Software: DeepStream Middleware: TF, TensorRT Hardware: Nvidia Xavier Configuration: Default number of counters number of splitters latency (ms) 100 150 1 200 250 2 300 Cubic Interpolation Over Finer Grid 2 4 3 6 8 4 10 12 5 14 16 6 18 Budget Exhausted? Yes No 5- Update Causal Performance Model Query Engine 4- Estimate Causal Queries Estimate probability of satisfying QoS if BufferSize is set to 6k? 2- Learn Causal Performance Model Performance Debugging Performance Optimization 3- Translate Perf. Query to Causal Queries •What is the root-cause of observed perf. fault? •How do I fix the misconfig.? •How can I improve throughput without sacrificing accuracy? •How do I understand perf behavior? Measure performance of the configuration(s) that maximizes information gain Performance Data Causal Model P(Th > 40/s|do(Buffersize = 6k)) 1- Specify Performance Query QoS : Th > 40/s Observed : Th < 30/s ± 5/s

I played a very minor role

Arti fi cial Intelligence and Systems Laboratory (AISys Lab) Machine Learning Computer Systems Autonomy AI/ML Systems https://pooyanjamshidi.github.io/AISys/ 97 Ying Meng (PhD student) Shuge Lei (PhD student) Kimia Noorbakhsh (Undergrad) Shahriar Iqbal (PhD student) Jianhai Su (PhD student) M.A. Javidian (postdoc) Sponsors, thanks! Fatemeh Ghofrani (PhD student) Abir Hossen (PhD student) Hamed Damirchi (PhD student) Mahdi Shari fi (PhD student) Lane Stanley (Intern)

98 Rahul Krishna Columbia Shahriar Iqbal UofSC M. A. Javidian Purdue Baishakhi Ray Columbia Christian Kästner CMU Sven Apel Saarland Marco Valtorta UofSC Madelyn Khoury REU student Forest Agostinelli UofSC Causal AI for Systems Causal AI for Robot Learning (Causal RL + Transfer Learning + Robotics) Abir Hossen UofSC Theory of Causal AI Ahana Biswas IIT Om Pandey KIIT Hamed Damirchi UofSC Causal AI for Adversarial ML Ying Meng UofSC Fatemeh Ghofrani UofSC Mahdi Shari fi UofSC Collaborators (Causal AI) Sugato Basu Google AdsAI Garima Pruthi Google AdsAI Causal Representation Learning

No content