https://songkou.github.io/tags/perftest/Recent content in Perftest on Council of ElrondHugoenSat, 18 Jul 2026 20:00:00 +0800https://songkou.github.io/posts/perftest/Sat, 18 Jul 2026 20:00:00 +0800https://songkou.github.io/posts/perftest/<h2 id="what-is-perftest">What is perftest?</h2> <p><strong>perftest</strong> is the standard micro-benchmark suite for RDMA — it validates and benchmarks InfiniBand, hardware RoCE, and Soft-RoCE links, not ordinary TCP networking. Each tool measures one RDMA verb:</p> <table> <thead> <tr> <th>Test</th> <th>Measures</th> <th>Real-world workloads it models</th> </tr> </thead> <tbody> <tr> <td><code>ib_write_bw</code></td> <td>RDMA Write bandwidth</td> <td>Data replication, storage writes, GPU/bulk data transfers</td> </tr> <tr> <td><code>ib_write_lat</code></td> <td>RDMA Write latency</td> <td>Updating remote memory — small one-sided writes</td> </tr> <tr> <td><code>ib_read_bw</code></td> <td>RDMA Read bandwidth</td> <td>Distributed storage and database reads, remote memory access</td> </tr> <tr> <td><code>ib_read_lat</code></td> <td>RDMA Read latency</td> <td>Fetching remote memory — storage/database-style lookups</td> </tr> <tr> <td><code>ib_send_bw</code></td> <td>Send/Receive bandwidth</td> <td>MPI messages, RPC, request queues</td> </tr> <tr> <td><code>ib_send_lat</code></td> <td>Send/Receive latency</td> <td>Small request/response round trips — RPC and MPI latency</td> </tr> <tr> <td><code>ib_atomic_bw</code></td> <td>Atomic operation rate</td> <td>Distributed locks and counters</td> </tr> <tr> <td><code>ib_atomic_lat</code></td> <td>Atomic latency</td> <td>Synchronization primitives</td> </tr> </tbody> </table> <h3 id="write-read-and-send-are-different-operations">Write, Read, and Send are different operations</h3> <p>The grouping in the table is not accidental — the three verbs have different semantics:</p>