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
con fi gurable 10 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 can cause performance to take abnormally large values • Faulty con fi gurations take the tail values (worse than 99.99th percentile) • Certain con fi gurations can cause faults on multiple performance objectives. 13
between software and hardwar e • These can result in non-functional fault s ◦ Affecting non-functional system properties like latency, throughput, energy consumption, etc. 14 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
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 cod e 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
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? ?
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 26 Default Con fi guration Optimal Con fi guration better better • Default is ba d • 2X-10X faster than worst • Noticeably faster than median
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. 32
increases Latency More GPU memory usage should reduce latency not increase it. Counterintuitive! Any ML-/statistical models built on this data will be incorrect !
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.
interacting variables as a causal graph 36 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
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
TF, TensorR T Hardware: Nvidia Xavie r 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 Performanc e Debugging Performanc e 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
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 Performanc e Debugging Performanc e 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
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 Performanc e Debugging Performanc e 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
× 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-functiona l measurable/quanti fi able aspect
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)
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
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 Performanc e Debugging Performanc e 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
non-functional faults Objective Causal Debugging: An example of downstream performance task Ὂ Use causal models to model various cross-stack configuration interactions; an d Ὂ Counterfactual reasoning to recommend fixes for these misconfigurations Approach
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 updat e observationa l 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)
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 updat e observationa l data About 25 sample configurations (training data)
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
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.
may be too many causal path s • We need to select the most useful one s • 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.
the ACE of all pairs of adjacent nodes in the pat h • Rank paths from highest path ACE (PACE) score to the lowes t • 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.
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 updat e observationa l data About 25 sample configurations (training data)
“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 G b • Swap Memory was initially set to 2 Gb • We observed high latency when Swap was set to 2 G b • 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?
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
^ 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
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. !
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 Performanc e Debugging Performanc e 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
Buffer Size Batch Size Enabl e Padding 1- Evaluate Candidate Interventions FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enabl e 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
Buffer Size Batch Size Enabl e Padding 1- Evaluate Candidate Interventions FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enabl e 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
Buffer Size Batch Size Enabl e Padding 1- Evaluate Candidate Interventions FPS Energy Branch Misses Cache Misses No of Cycles Bitrate Buffer Size Batch Size Enabl e 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
“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
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
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
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.
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 cod e 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
Cores ✓ ✓ ✓ CPU Freq. ✓ ✓ ✓ EMC Freq. ✓ ✓ ✓ GPU Freq. ✓ ✓ ✓ Sched. Policy ✓ Sched. Runtime ✓ Sched. Child Proc ✓ Dirty Bg. Ratio ✓ Drop Caches ✓ CUDA_STATIC_R T ✓ ✓ ✓ Swap Memory ✓ CADET 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 accuratel y X No unnecessary change s X Better improvements than forum’s recommendatio n X Much faster Results The user expected 30-40% gain
benchmark datase t ◦ 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 dissipatio n ◦ Repeat 5x and average 82 Multiple Faults ! Latency Faults ! Energy Faults !
Manually investigate the root-cause ◦ “Fix” the misconfigurations • A “fix” implies the configuration no longer has tail performanc e ◦ User defined benchmark (i.e., 10th percentile) ◦ Or some QoS/SLA benchmark • Record the configurations that were changed 83 Multiple Faults ! Latency Faults ! Energy Faults !
NFP fault − NFP repair NFP fault × 100 Repair Quality NFP = Non-Functional Propert y (e.g., Latency, Energy, etc.) Repair value Faulty value Larger the gain, better the repair
CADET perform compared to Model based Diagnostics X Finds the root-causes accurately X Better gain X Much faster Takeaways More accurate tha n ML-based methods Better Gain Up to 20x faster
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 92
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. 93
Model for di ff erent downstream systems tasks 2. The learned causal model is transferable across di ff erent environments 94 Software: DeepStream Middleware: TF, TensorR T Hardware: Nvidia Xavie r 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 Performanc e Debugging Performanc e 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
pjamshidi
monochromegane
hotekagi
beomi
asonas
yumulab
sansan_randd
takase
takumi1001
masatoto
abap34
sgk
a1da4
marktimemedia
chrisshort
ammeep
rmw
kastner
andyhume
danielanewman
productmarketing
scottboms
lauravandoore
eileencodes
tanoku
