At the moment, we’re exploring how Ethernet stacks up towards InfiniBand in AI/ML environments, specializing in how Cisco Silicon One™ manages community congestion and enhances efficiency for AI/ML workloads. This submit emphasizes the significance of benchmarking and KPI metrics in evaluating community options, showcasing the Cisco Zeus Cluster geared up with 128 NVIDIA® H100 GPUs and cutting-edge congestion administration applied sciences like dynamic load balancing and packet spray.
Networking requirements to satisfy 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 site visitors throughout mannequin coaching calls for a low-latency, lossless community cloth. InfiniBand has been a dominant know-how within the high-performance computing (HPC) setting and recently 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 cloth options for AI/ML
Community site visitors patterns range based mostly on mannequin dimension, structure, and parallelization methods utilized in accelerated coaching. To guage AI/ML community cloth 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 totally 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 normal 3-stage leaf-spine Clos cloth 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 software throughput throughout coaching and inference communication part amongst GPUs
- MLCommons MLPerf set of benchmarks, which evaluates essentially 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 = Specific Congestion Notification and Precedence Move Management
NCCL benchmarking towards congestion avoidance options
Congestion builds up in the course of the again propagation stage of the coaching course of, the place a gradient sync is required amongst all of the GPUs collaborating in coaching. Because the mannequin dimension will increase, so does the gradient dimension and the variety of GPUs. This creates huge micro-congestion within the community cloth. Determine 2 exhibits outcomes of the JCT and site visitors distribution benchmarking. Be aware how Cisco Silicon One helps a set of superior options for congestion avoidance, reminiscent of dynamic load balancing (DLB) and packet spray methods, and Knowledge Heart Quantized Congestion Notification (DCQCN) for congestion administration.
Determine 2 illustrates how the NCCL benchmarks stack up towards totally different congestion avoidance options. We examined the commonest collectives with a number of totally 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 essentially the most congestion because of the nature of communication. Though JCT is essentially the most understood metric from an software’s perspective, JCT doesn’t present how successfully the community is utilized—one thing the infrastructure group must know. This information might assist them to:
- Enhance the community utilization to get higher JCT
- Know what number of workloads can share the community cloth with out adversely impacting JCT
- Plan for capability as use instances improve
To gauge community cloth utilization, we calculated Jain’s Equity Index, the place LinkTxᵢ is the quantity of transmitted site visitors on cloth hyperlink:
The index worth ranges from 0.0 to 1.0, with greater values being higher. A worth of 1.0 represents the right distribution. The Site visitors Distribution on Cloth Hyperlinks chart in Determine 2 exhibits how DLB and packet spray algorithms create a near-perfect Jain’s Equity Index, so site visitors distribution throughout the community cloth is sort of excellent. ECMP makes use of static hashing, and relying on stream entropy, it could actually result in site visitors polarization, inflicting micro-congestion and negatively affecting JCT.
Silicon One versus NVIDIA Spectrum-X and InfiniBand
The NCCL Benchmark – Aggressive Evaluation (Determine 3) exhibits how Cisco Silicon One performs towards NVIDIA Spectrum-X and InfiniBand applied sciences. The information for NVIDIA was taken from the SemiAnalysis publication. Be aware that Cisco doesn’t understand how these exams have been carried out, however we do know that the cluster dimension and GPU to community cloth 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 particular mathematical equation reported throughout benchmarking. Determine 3 exhibits that Cisco Silicon One – All Scale back performs comparably to NVIDIA Spectrum-X and InfiniBand throughout numerous message sizes.
Community cloth efficiency evaluation
The IBPerf Benchmark compares RDMA efficiency towards ECMP, DLB, and packet spray, that are essential for assessing community cloth efficiency. Incast eventualities, the place a number of GPUs ship information to 1 GPU, typically trigger congestion. We simulated these situations utilizing IBPerf instruments.
Determine 4 exhibits how Aggregated Session Throughput and JCT reply to totally different congestion avoidance algorithms: ECMP, DLB, and packet spray. DLB and packet spray attain Hyperlink Bandwidth, bettering JCT. It additionally illustrates how DCQCN handles micro-congestions, with PFC and ECN ratios bettering 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 perfect site visitors distribution.
Coaching and inference benchmark
The MLPerf Benchmark – Coaching and Inference, revealed by the MLCommons group, goals to allow truthful comparability of AI/ML methods and options.
We targeted on AI/ML information heart options by executing coaching and inference benchmarks. To attain optimum outcomes, we extensively tuned throughout compute, storage, and networking parts utilizing congestion administration options of Cisco Silicon One. Determine 5 exhibits comparable efficiency throughout numerous 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 progressive developments showcase the unwavering dedication of Cisco to supply sturdy, high-performance networking options that meet the rigorous calls for of at this time’s AI/ML functions.
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 you should develop, deploy, and handle AI workloads securely while you select Cisco Safe AI Manufacturing unit with NVIDIA.
Share:
