1 / 42

A CASE STUDY

A CASE STUDY. Section 3. Overview In this section we will show the negative consequences of the lack of coordination in a large scale machine We analyze the behavior of a complex scientific application representative of the ASCI workload on a large scale supercomputer

blake-ramirez
Télécharger la présentation

A CASE STUDY

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. A CASE STUDY

  2. Section 3 • Overview • In this section we will show the negative consequences of the lack of coordination in a large scale machine • We analyze the behavior of a complex scientific application representative of the ASCI workload on a large scale supercomputer • A case study that emphasizes the importance of the coordination in the network and in the system software

  3. ASCI Q • 2,048 ES45 Alphaservers, with 4 processors/node • 16 GB of memory per node • 8,192 processors in total • 2 independent network rails, Quadrics Elan3 • > 8192 cables • 20 Tflops peak, #2 in the top 500 lists • A complex human artifact

  4. Dealing with the complexity of a real system • In this section of the tutorial we provide insight into our methodology, that we used to substantially improve the performance of ASCI Q. • This methodology is based on an arsenal of • analytical models • custom microbenchmarks • full applications • discrete event simulators • Dealing with the complexity of the machine and the complexity of a real parallel application, SAGE, with > 150,000 lines of Fortran & MPI code

  5. Overview • Our performance expectations for ASCI Q and the reality • Identification of performance factors • Application performance and breakdown into components • Detailed examination of system effects • A methodology to identify operating systems effects • Effect of scaling – up to 2000 nodes/ 8000 processors • Quantification of the impact • Towards the elimination of overheads • demonstrated over 2x performance improvement • Generalization of our results: application resonance • Bottom line: the importance of the integration of the various system across nodes

  6. Performance of SAGE on 1024 nodes There is a difference why ? • Performance consistent across QA and QB (the two segments of ASCI Q, with 1024 nodes/4096 processors each) • Measured time 2x greater than model (4096 PEs) Lower is better!

  7. Using fewer PEs per Node Test performance using 1,2,3 and 4 PEs per node Lower is better!

  8. Using fewer PEs per node (2) Measurements match model almost exactly for 1,2 and 3 PEs per node! Performance issue only occurs when using 4 PEs per node

  9. Mystery #1 SAGE performs significantly worse on ASCI Q than was predicted by our model

  10. SAGE performance components • Look at SAGE in terms of main components: • Put/Get (point-to-point boundary exchange) • Collectives (allreduce, broadcast, reduction) Performance issue seems to occur only on collective operations

  11. Performance of the collectives Allreduce Latency 3 1 process per node 2 processes per node 2.5 3 processes per node 2 Latency ms 1.5 1 0.5 0 0 100 200 300 400 500 600 700 800 900 1000 Nodes • Measure collective performance separately 4 processes per node • Collectives (e.g., allreduce and barrier) mirror the performance of the application

  12. Identifying the problem within Sage Sage Simplify Allreduce

  13. Exposing the problems with simple benchmarks Allreduce Challenge: identify the simplest benchmark that exposes the problem Add complexity Benchmarks

  14. Interconnection network and communication libraries • The initial (obvious) suspects were the interconnection network and the MPI implementation • We tested in depth the network, the low level transmission protocols and several allreduce algorithms • We also implemented allreduce in the Network Interface Card • By changing the synchronization mechanism we were able to reduce the latency of an allreduce benchmark by a factor of 7 • But we only got small improvements in Sage (5%)

  15. Mystery #2 Although SAGE spends half of its time in allreduce (at 4,096 processors), making allreduce 7 times faster leads to a small performance improvement

  16. Computational noise • After having ruled out the network and MPI we focused our attention on the compute nodes • Our hypothesis is that the computational noise is generated inside the processing nodes • This noise “freezes” a running process for a certain amount of time and generates a “computational” hole

  17. Computational noise: intuition • The computation of one process is interrupted by an external event (e.g., system daemon or kernel) P P P P 2 0 1 3 • Running 4 processes on all 4 processors of an Alphaserver ES45

  18. Computational noise: 3 processes on 3 processors • The “noise” can run on the 4th processor without interrupting the other 3 processes P P P 2 0 1 • Running 3 processes on 3 processors of an Alphaserver ES45 IDLE

  19. Coarse grained measurement • We execute a computational loop for 1,000 seconds on all 4,096 processors of QB START END P 1 P 2 P 3 P 4 TIME

  20. Coarse grained computational overhead per process • The slowdown per process is small, between 1% and 2.5% lower is better

  21. Mystery #3 Although the “noise” hypothesis could explain SAGE’s suboptimal performance, the microbenchmarks of per-processor noise indicate that at most 2.5% of performance is lost to noise

  22. Fine grained measurement • We run the same benchmark for 1000 seconds, but we measure the run time every millisecond • Fine granularity representative of many ASCI codes

  23. Fine grained computational overhead per node • We now compute the slowdown per-node, rather than per-process • The noise has a clear, per cluster, structure Optimum is 0 (lower is better)

  24. Finding #1 Analyzing noise on a per-node basis reveals a regular structure across nodes

  25. Noise in a 32 Node Cluster • The Q machine is organized in 32 node clusters (TruCluster) • In each cluster there is a cluster manager (node 0), a quorum node (node 1) and the RMS data collection (node 31)

  26. Per node noise distribution • Plot distribution of one million, 1 ms computational chunks • In an ideal, noiseless, machine the distribution graph is • a single bar at 1 ms of 1 million points per process (4 million per node) • Every outlier identifies a computation that was delayed by external interference • We show the distributions for the standard cluster node, and also nodes 0, 1 and 31

  27. Cluster Node (2-30) • 10% of the times the execution of the 1 ms chunk of computation is delayed

  28. Node 0, Cluster Manager • We can identify 4 main sources of noise

  29. Node 1, Quorum Node • One source of heavyweight noise (335 ms!)

  30. Node 31 • Many fine grained interruptions, between 6 and 8 milliseconds

  31. The effect of the noise . . . . . . . . . . . . . . . . . . . . . . . . . . . • An application is usually a sequence of a computation followed by a synchronization (collective): • But if an event happens on a single node then it can affect all the other nodes

  32. Effect of System Size . . . . . . . . . . . . • The probability of a random event occurring increases with the node count.

  33. Tolerating Noise: Buffered Coscheduling (BCS) . . . . . . . . . . . . . . . . . . . . . . . . We can tolerate the noise by coscheduling the activities of the system software on each node

  34. Discrete Event Simulator:used to model noise • DES used to examine and identify impact of noise: takes as input the harmonics that characterize the noise • Noise model closely approximates experimental data • The primary bottleneck is the fine-grained noise generated by the compute nodes (Tru64) Lower is better

  35. Finding #2 On fine-grained applications, more performance is lost to short but frequent noise on all nodes than to long but less frequent noise on just a few nodes

  36. Incremental noise reduction • removed about 10 daemons from all nodes (including: envmod, insightd, snmpd, lpd, niff) • decreased RMS monitoring frequency by a factor of 2 on each node (from an interval of 30s to 60s) • moved several daemons from nodes 1 and 2 to node 0 on each cluster.

  37. Improvements in the Barrier Synchronization Latency

  38. Resulting SAGE Performance • Nodes 0 and 31 also configured out in the optimization

  39. Finding #3 We were able to double SAGE’s performance by selectively removing noise caused by several types of system activities

  40. Generalizing our results:application resonance • The computational granularity of a balanced bulk-synchronous application correlates to the type of noise. • Intuition: • any noise source has a negative impact, a few noise sources tend to have a major impact on a given application. • Rule of thumb: • the computational granularity of the application “enters in resonance” with the noise of the same order of magnitude • The performance can be enhanced by selectively removing sources of noise • We can provide a reasonable estimate of the performance improvement knowing the computational granularity of a given application.

  41. Cumulative Noise Distribution, Sequence of Barriers with No Computation • Most of the latency is generated by the fine-grained, high-frequency noisie of the cluster nodes

  42. Conclusions • Combination of Measurement, Simulation and Modeling to • Identify and resolve performance issues on Q • Used modeling to determine that a problem exists • Developed computation kernels to quantify O/S events: • Effect increases with the number of nodes • Impact is determined by the computation granularity in an application • Application performance has significantly improved • Method also being applied to other large-systems

More Related