As we speak, we’re exploring how Ethernet stacks up in opposition to InfiniBand in AI/ML environments, specializing in how Cisco Silicon One™ manages community congestion and enhances efficiency for AI/ML workloads. This put up emphasizes the significance of benchmarking and KPI metrics in evaluating community options, showcasing the Cisco Zeus Cluster outfitted with 128 NVIDIA® H100 GPUs and cutting-edge congestion administration applied sciences like dynamic load balancing and packet spray.
Networking requirements to fulfill the wants of AI/ML workloads
AI/ML coaching workloads generate repetitive micro-congestion, stressing community buffers considerably. The east-to-west GPU-to-GPU visitors throughout mannequin coaching calls for a low-latency, lossless community material. InfiniBand has been a dominant expertise within the high-performance computing (HPC) setting and currently within the AI/ML setting.
Ethernet is a mature various, with superior options that may handle the rigorous calls for of the AI/ML coaching workloads and Cisco Silicon One can successfully execute load balancing and handle congestion. We got down to benchmark and evaluate Cisco Silicon One versus NVIDIA Spectrum-X™ and InfiniBand.
Analysis of community material options for AI/ML
Community visitors patterns fluctuate based mostly on mannequin dimension, structure, and parallelization methods utilized in accelerated coaching. To guage AI/ML community material options, we recognized related benchmarks and key efficiency indicator (KPI) metrics for each AI/ML workload and infrastructure groups, as a result of they view efficiency by means of completely different lenses.
We established complete exams to measure efficiency and generate metrics particular to AI/ML workload and infrastructure groups. For these exams, we used the Zeus Cluster, that includes devoted backend and storage with a typical 3-stage leaf-spine Clos material community, constructed with Cisco Silicon One–based mostly platforms and 128 NVIDIA H100 GPUs. (See Determine 1.)
We developed benchmarking suites utilizing open-source and industry-standard instruments contributed by NVIDIA and others. Our benchmarking suites included the next (see additionally Desk 1):
- Distant Direct Reminiscence Entry (RDMA) benchmarks—constructed utilizing IBPerf utilities—to guage community efficiency throughout congestion created by incast
- NVIDIA Collective Communication Library (NCCL) benchmarks, which consider utility throughput throughout coaching and inference communication part amongst GPUs
- MLCommons MLPerf set of benchmarks, which evaluates probably the most understood metrics, job completion time (JCT) and tokens per second by the workload groups
Legend:
JCT = Job Completion Time
Bus BW = Bus bandwidth
ECN/PFC = Express Congestion Notification and Precedence Stream Management
NCCL benchmarking in opposition to congestion avoidance options
Congestion builds up throughout the again propagation stage of the coaching course of, the place a gradient sync is required amongst all of the GPUs taking part in coaching. Because the mannequin dimension will increase, so does the gradient dimension and the variety of GPUs. This creates large micro-congestion within the community material. Determine 2 reveals outcomes of the JCT and visitors distribution benchmarking. Word how Cisco Silicon One helps a set of superior options for congestion avoidance, equivalent to dynamic load balancing (DLB) and packet spray methods, and Knowledge Middle Quantized Congestion Notification (DCQCN) for congestion administration.
Determine 2 illustrates how the NCCL benchmarks stack up in opposition to completely different congestion avoidance options. We examined the commonest collectives with a number of completely different message sizes to focus on these metrics. The outcomes present that JCT improves with DLB and packet spray for All-to-All, which causes probably the most congestion because of the nature of communication. Though JCT is probably the most understood metric from an utility’s perspective, JCT doesn’t present how successfully the community is utilized—one thing the infrastructure crew must know. This data may assist them to:
- Enhance the community utilization to get higher JCT
- Know what number of workloads can share the community material with out adversely impacting JCT
- Plan for capability as use circumstances enhance
To gauge community material utilization, we calculated Jain’s Equity Index, the place LinkTxᵢ is the quantity of transmitted visitors on material hyperlink:
The index worth ranges from 0.0 to 1.0, with increased values being higher. A worth of 1.0 represents the proper distribution. The Visitors Distribution on Material Hyperlinks chart in Determine 2 reveals how DLB and packet spray algorithms create a near-perfect Jain’s Equity Index, so visitors distribution throughout the community material is sort of excellent. ECMP makes use of static hashing, and relying on circulation entropy, it could actually result in visitors polarization, inflicting micro-congestion and negatively affecting JCT.
Silicon One versus NVIDIA Spectrum-X and InfiniBand
The NCCL Benchmark – Aggressive Evaluation (Determine 3) reveals how Cisco Silicon One performs in opposition to NVIDIA Spectrum-X and InfiniBand applied sciences. The information for NVIDIA was taken from the SemiAnalysis publication. Word that Cisco doesn’t understand how these exams had been carried out, however we do know that the cluster dimension and GPU to community material connectivity is just like the Cisco Zeus Cluster.
Bus Bandwidth (Bus BW) benchmarks the efficiency of collective communication by measuring the pace of operations involving a number of GPUs. Every collective has a selected mathematical equation reported throughout benchmarking. Determine 3 reveals that Cisco Silicon One – All Cut back performs comparably to NVIDIA Spectrum-X and InfiniBand throughout varied message sizes.
Community material efficiency evaluation
The IBPerf Benchmark compares RDMA efficiency in opposition to ECMP, DLB, and packet spray, that are essential for assessing community material efficiency. Incast situations, the place a number of GPUs ship knowledge to at least one GPU, usually trigger congestion. We simulated these circumstances utilizing IBPerf instruments.
Determine 4 reveals how Aggregated Session Throughput and JCT reply to completely different congestion avoidance algorithms: ECMP, DLB, and packet spray. DLB and packet spray attain Hyperlink Bandwidth, enhancing JCT. It additionally illustrates how DCQCN handles micro-congestions, with PFC and ECN ratios enhancing with DLB and considerably dropping with packet spray. Though JCT improves barely from DLB to packet spray, the ECN ratio drops dramatically on account of packet spray’s excellent visitors distribution.
Coaching and inference benchmark
The MLPerf Benchmark – Coaching and Inference, revealed by the MLCommons group, goals to allow honest comparability of AI/ML techniques and options.
We targeted on AI/ML knowledge middle options by executing coaching and inference benchmarks. To attain optimum outcomes, we extensively tuned throughout compute, storage, and networking elements utilizing congestion administration options of Cisco Silicon One. Determine 5 reveals comparable efficiency throughout varied platform distributors. Cisco Silicon One with Ethernet performs like different vendor options for Ethernet.
Conclusion
Our deep dive into Ethernet and InfiniBand inside AI/ML environments highlights the outstanding prowess of Cisco Silicon One in tackling congestion and boosting efficiency. These revolutionary developments showcase the unwavering dedication of Cisco to offer strong, high-performance networking options that meet the rigorous calls for of in the present day’s AI/ML purposes.
Many due to Vijay Tapaskar, Will Eatherton, and Kevin Wollenweber for his or her help on this benchmarking course of.
Discover safe AI infrastructure
Uncover safe, scalable, and high-performance AI infrastructure it’s essential develop, deploy, and handle AI workloads securely once you select Cisco Safe AI Manufacturing facility with NVIDIA.
Share:
