Implementation & Comparison Of Rdma Over Ethernet
Download as PPTX, PDF0 likes1,590 views
The document reports on a study evaluating the performance of RDMA over Converged Ethernet (RoCE) compared to traditional Infiniband and software RoCE implementations. Results demonstrate that RoCE can achieve significant performance gains in both latency and bandwidth, closely approaching Infiniband capabilities, especially for specific message sizes. Further investigation is suggested to optimize RoCE configurations and assess performance under different conditions.
Implementation & Comparison Of Rdma Over Ethernet
- 1. LA-UR 10-05188Implementation & Comparison of RDMA Over EthernetStudents: Lee Gaiser, Brian Kraus, and James WernickeMentors: Andree Jacobson, Susan Coulter, JharrodLaFon, and Ben McClelland
- 2. SummaryBackgroundObjectiveTesting EnvironmentMethodologyResultsConclusionFurther WorkChallengesLessons LearnedAcknowledgmentsReferences & LinksQuestions
- 3. Background : Remote Direct Memory Access (RDMA)RDMA provides high-throughput, low-latency networking:Reduce consumption of CPU cyclesReduce communication latencyImages courtesy of http://www.hpcwire.com/features/17888274.html
- 4. Background : InfiniBandInfiniband is a switched fabric communication link designed for HPC:High throughputLow latencyQuality of serviceFailoverScalabilityReliable transportHow do we interface this high performance link with existing Ethernet infrastructure?
- 5. Background : RDMA over Converged Ethernet (RoCE)Provide Infiniband-like performance and efficiency to ubiquitous Ethernet infrastructure.Utilize the same transport and network layers from IB stack and swap the link layer for Ethernet.Implement IB verbs over Ethernet.Not quite IB strength, but it’s getting close.As of OFED 1.5.1, code written for OFED RDMA auto-magically works with RoCE.
- 6. ObjectiveWe would like to answer the following questions:What kind of performance can we get out of RoCE on our cluster?Can we implement RoCE in software (Soft RoCE) and how does it compare with hardware RoCE?
- 7. Testing EnvironmentHardware:HP ProLiant DL160 G6 serversMellanox MNPH29B-XTC 10GbE adapters50/125 OFNR cablingOperating System:CentOS 5.32.6.32.16 kernelSoftware/Drivers:Open Fabrics Enterprise Distribution (OFED) 1.5.2-rc2 (RoCE) & 1.5.1-rxe (Soft RoCE)OSU Micro Benchmarks (OMB) 3.1.1OpenMPI 1.4.2
- 8. MethodologySet up a pair of nodes for each technology:IB, RoCE, Soft RoCE, and no RDMAInstall, configure & run minimal services on test nodes to maximize machine performance.Directly connect nodes to maximize network performance.Acquire latency benchmarksOSU MPI Latency TestAcquire bandwidth benchmarksOSU MPI Uni-Directional Bandwidth TestOSU MPI Bi-Directional Bandwidth TestScript it all to perform many repetitions
- 10. Results : Uni-directional Bandwidth
- 11. Results : Bi-directional Bandwidth
- 12. Results : AnalysisRoCE performance gains over 10GbE:
- 13. Up to 5.7x speedup in latency
- 14. Up to 3.7x increase in bandwidth
- 15. IB QDR vs. RoCE:
- 16. IB less than 1µs faster than RoCE at 128-byte message.
- 17. IB peak bandwidth is 2-2.5x greater than RoCE.ConclusionRoCE is capable of providing near-Infiniband QDR performance for:Latency-critical applications at message sizes from 128B to 8KBBandwidth-intensive applications for messages <1KB.Soft RoCE is comparable to hardware RoCE at message sizes above 65KB.Soft RoCE can improve performance where RoCE-enabled hardware is unavailable.
- 18. Further Work & QuestionsHow does RoCE perform over collectives?Can we further optimize RoCE configuration to yield better performance?Can we stabilize the Soft RoCE configuration?How much does Soft RoCE affect the compute nodes ability to perform?How does RoCE compare with iWARP?
- 19. ChallengesFinding an OS that works with OFED & RDMA:Fedora 13 was too new.Ubuntu 10 wasn’t supported.CentOS 5.5 was missing some drivers.Had to compile a new kernel with IB/RoCE support.Built OpenMPI 1.4.2 from source, but wasn’t configured for RDMA; used OpenMPI 1.4.1 supplied with OFED instead.The machines communicating via Soft RoCE frequently lock up during OSU bandwidth tests.
- 20. Lessons LearnedInstalling and configuring HPC clustersBuilding, installing, and fixing Linux kernel, modules, and driversWorking with IB, 10GbE, and RDMA technologiesUsing tools such as OMB-3.1.1 and netperf for benchmarking performance
- 21. AcknowledgmentsAndree JacobsonSusan CoulterJharrodLaFonBen McClellandSam GutierrezBob Pearson
- 22. Questions?
- 23. References & LinksSubmaroni, H. et al. RDMA over Ethernet – A Preliminary Study. OSU. http://nowlab.cse.ohio-state.edu/publications/conf-presentations/2009/subramoni-hpidc09.pdfFeldman, M. RoCE: An Ethernet-InfiniBand Love Story. HPCWire.com. April 22, 2010.http://www.hpcwire.com/blogs/RoCE-An-Ethernet-InfiniBand-Love-Story-91866499.htmlWoodruff, R. Access to InfiniBand from Linux.Intel. October 29, 2009. http://software.intel.com/en-us/articles/access-to-infiniband-from-linux/OFED 1.5.2-rc2http://www.openfabrics.org/downloads/OFED/ofed-1.5.2/OFED-1.5.2-rc2.tgzOFED 1.5.1-rxehttp://www.systemfabricworks.com/pub/OFED-1.5.1-rxe.tgzOMB 3.1.1http://mvapich.cse.ohio-state.edu/benchmarks/OMB-3.1.1.tgz
Editor's Notes
- #2: Introduce yourself, the institute, your teammates, and what you’ve been working on for the past two months.
- #3: Rephrase these bullet points with a little more elaboration.
- #4: Emphasize how RDMA eliminates unnecessary communication.
- #5: Explain that we are using IB QDR and what that means.
- #6: Emphasize that the biggest advantage of RoCE is latency, not necessarily bandwidth. Talk about 40Gb & 100Gb Ethernet on the horizon.
- #9: OSU benchmarks were more appropriate than netperf
- #10: The highlight here is that latency between IB & RoCE differs by only 1.7us at 128 byte messages. It continues to be very close up through 4K messages. Also notice that latency in RoCE and no RDMA converge at higher messages.
- #11: Note that RoCE & IB are very close up to 1K message size. IB QDR peaks out at 3MB/s, RoCE at 1.2MB/s
- #12: Note that the bandwidth trends are similar to the uni-directional bandwidths. Explain that Soft RoCE could not complete this test. IB QDR peaks at 5.5 MB/s and RoCE peaks at 2.3 MB/s.