IP Storage and Big Data

IP storage workloads (NVMe-oF, iSCSI, and object storage protocols) place unusual demands on the underlying network fabric: lossless delivery within storage clusters to avoid protocol-level retransmit storms, deep packet buffers to absorb traffic bursts from parallel reads and writes across large numbers of drives, and telemetry granularity sufficient to distinguish storage-tier congestion from application-tier congestion during performance investigations. Arista's R-series platforms provide the buffer depth and telemetry depth for these workloads, with the 7500R3 delivering up to 16 GB of packet buffer per line card and EOS streaming telemetry capturing per-queue depth and congestion events at millisecond granularity.

Big data analytics frameworks such as Spark and Hadoop generate east-west traffic patterns characterized by large shuffle transfers between compute nodes that complete within a relatively narrow time window, a pattern that creates simultaneous incast traffic from many sources to a smaller number of receivers. Deep buffering prevents the packet drops that cause TCP to back off and extend the shuffle phase, keeping the compute cluster at full utilization through the data movement phases that otherwise become the throughput bottlenecks limiting job completion time.

Arista's lossless fabric capabilities extend to RoCE (RDMA over Converged Ethernet) workloads that require Priority Flow Control (PFC) and Explicit Congestion Notification (ECN) for loss-free delivery at line rate. For AI/ML training workloads where AllReduce collective operations generate simultaneous RDMA traffic from hundreds of GPU nodes, a lossless EOS fabric ensures that a single congested link does not retransmit the entire collective communication and stall the training iteration, a constraint that becomes increasingly critical as GPU cluster sizes grow beyond thousands of nodes.