Apache Iggy Bench CLI
The interactive Bench CLI allows you to perform various benchmarking on the Apache Iggy server.
Iggy Bench CLI can be installed with cargo install iggy-bench and then simply accessed by typing iggy-bench in your terminal.
The WebSocket transport command is websocket, with ws as a shorthand.
Producer and consumer counts default to six. Pinned workloads also default to six streams. Override the counts for the workload and available CPUs.
Examples and topic options
Start Iggy before running benchmarks. The tool connects to an existing server. Run iggy-bench examples for all eight benchmark kinds, their aliases, all four transports, and topic-option combinations. Global options precede the benchmark kind, kind-specific options follow it, and the server address follows the transport.
Both --durability and --consumer-offset-durability independently default to replicated. The following runs use the same workload with different topic policies:
# Both policies replicated, the default. iggy-bench balanced-producer-and-consumer-group tcp # Persisted message acknowledgments with replicated offsets. iggy-bench --durability persisted balanced-producer-and-consumer-group tcp # Replicated message acknowledgments with persisted offsets. iggy-bench --consumer-offset-durability persisted balanced-producer-and-consumer-group tcp # Both policies persisted. iggy-bench --durability persisted --consumer-offset-durability persisted balanced-producer-and-consumer-group tcp
These options apply when the benchmark creates topics. They do not change existing topics with --reuse-streams. Both policies normally write data to disk. persisted adds a stable-storage completion requirement. Consumer workloads using poll auto-commit still receive asynchronous poll responses, so this flag does not turn their poll latency into a measurement of acknowledged offset-store latency.
--messages-required-to-save controls a topic‘s segment-flush cadence independently of durability. --max-topic-size and --message-expiry are kind-specific topic options. Use the selected kind’s --help to inspect supported flags. OS writeback settings do not replace these policies.
Host preparation
Use a dedicated server and a separate load generator where possible. If they share a host, assign disjoint CPU sets and leave CPU capacity for kernel and network work. Match the deployment's CPU topology, storage, replication-group size, and durability policies.
| Area | Starting point | What to check |
|---|---|---|
| Runtime access | Permit the service to create io_uring instances. | Startup diagnostics, container syscall policy, and effective process limits. |
| File descriptors and locked memory | Size limits for connections, partitions, and runtime allocations. | /proc/PID/limits, not just the shell's ulimit. Set systemd service limits explicitly. |
| Swap | With enough RAM and disk-backed swap, compare vm.swappiness=10 against the existing value. | Swap activity, memory pressure, and p99 latency. This does not disable swap. |
| Huge pages | Start without an explicit reservation, then test a budgeted pool with a compatible allocator. | Actual page size, usage, NUMA placement, and memory left for page cache. |
| CPU placement | Use the server's [sharding] options within its allowed CPU set. | Per-core load, CPU steal time, cgroup throttling, and IRQ distribution. |
| Writeback | Start with the OS defaults. | Sustained disk latency and dirty-page buildup. Large percentage limits can mask short-run bottlenecks. |
| Networking | Start with the OS defaults. | Receive drops, retransmissions, connection backlogs, and bandwidth-delay requirements before increasing limits. |
Read the Linux tuning guide before reserving huge pages. vm.nr_hugepages=2048 reserves 4 GiB only with 2 MiB pages. MIMALLOC_RESERVE_HUGE_OS_PAGES=2048 requests 2048 one-GiB pages, or 2 TiB. They are different controls. The guide explains allocator configuration, THP, service limits, permissions, and reboot persistence.
Repeatable measurements
- Record the server and benchmark revisions, build profile, kernel, VM type, CPU allocation, allocator, memory limits, filesystem, and provisioned disk/network capacity.
- Keep workload shape, both durability policies, replication-group size, and topic options identical between compared runs. Use an explicit warmup and repeat runs.
- Observe
vmstat 1,iostat -xz 1,mpstat -P ALL 1, memory/I/O pressure, and network errors. Confirm the client has spare CPU and network capacity. - Run long enough to include segment flushes, checkpoints, and sustained device limits. Distinguish warm-cache tests, cold-cache tests, and storage throughput. Do not drop caches during a production workload.
- Change one host setting at a time, retain its previous value, and compare throughput, p50/p99/p99.9 latency, memory, and errors. Record successful changes in provisioning and recheck after reboot.
The output subcommand records results and accepts context for the run:
iggy-bench --warmup-time 10s --total-data 10GiB balanced-producer tcp output --identifier baseline --remark replicated-defaults
Choose --total-data for the duration and storage behavior being tested. This example is not a universal sufficient data volume.