Erlang VM Tuning

Riak runs on Erlang VM, so tuning VM parameters is part of performance optimization. The page maps Erlang flags and environment values to Riak configuration names, then explains which settings affect multiprocessing, scheduler behavior, port usage, buffering, storage limits, crash dumps, network tick checks, and shutdown timing.

Reading notes

#

• Riak is written almost entirely in Erlang and exposes only a subset of Erlang VM parameters through node configuration files.
• In Riak versions before 2.0, VM parameters were set in vm.args; in 2.0 and later, they are set in riak.conf, and vm.args values override riak.conf when both are present.
• SMP can be enabled, disabled, or set to auto; auto enables it only when the OS supports SMP and more than one logical processor is detected.
• The scheduler settings include total and online thread counts, wakeup interval, compaction of load, and utilization balancing.
• +sfwi is recommended at 500 and +scl at false when using the older vm.args system.
• Only one of compaction of load or utilization balancing can be active at a time; if both are false, Riak uses compaction of load.
• Port settings cover the Erlang distribution port range, the maximum number of concurrent ports or sockets, and the nodename used for cluster node identifiers.
• The asynchronous thread pool can be sized with erlang.async_threads, and each async thread has a configurable stack size.
• Kernel polling is on by default and can be turned off with erlang.K = off.
• Warning messages from error_logger can be mapped to warnings, errors, or info reports with erlang.W.
• erlang.process_limit sets the maximum number of system processes, and erlang.distribution_buffer_size controls the distribution buffer busy limit.
• erlang.max_ets_tables sets the ETS table limit, and higher values increase quick-access storage at the cost of RAM.
• erlang.crash_dump changes the crash dump location, and erlang.distribution.net_ticktime sets the interval for node liveness checks.
• erlang.shutdown_time sets how long the VM spends shutting down before killing existing processes.