From 97fa9fe68af5990a8ef7a71ccb7fc235d20a2c04 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Thu, 3 Jun 2010 12:43:57 +0000 Subject: OTP-8666 Enable automatic binding of schedulers when possible The runtime system will by default bind schedulers to logical processors using the default_bind bind type if the amount of schedulers are at least equal to the amount of logical processors configured, binding of schedulers is supported, and a CPU topology is available at startup. NOTE: If the Erlang runtime system is the only operating system process that binds threads to logical processors, this improves the performance of the runtime system. However, if other operating system processes (as for example another Erlang runtime system) also bind threads to logical processors, there might be a performance penalty instead. If this is the case you, are are advised to unbind the schedulers using the +sbtu command line argument, or by invoking erlang:system_flag(schedule r_bind_type, unbound). --- erts/doc/src/erl.xml | 36 ++++++++++++++++++++++++++++++++++++ erts/doc/src/erlang.xml | 19 +++++++++++++++---- erts/emulator/beam/erl_process.c | 38 ++++++++++++++++++++++++++++---------- 3 files changed, 79 insertions(+), 14 deletions(-) diff --git a/erts/doc/src/erl.xml b/erts/doc/src/erl.xml index bb741c7836..df80142ce1 100644 --- a/erts/doc/src/erl.xml +++ b/erts/doc/src/erl.xml @@ -41,6 +41,26 @@ to scroll back to text which has scrolled off the screen. The program must be used, however, in pipelines or if you want to redirect standard input or output.

+

As of ERTS version 5.8 (OTP-R14A) the runtime system will by + default bind schedulers to logical processors using the + default_bind bind type if the amount of schedulers are + at least equal to the amount of logical processors configured, + binding of schedulers is supported, and a CPU topology is + available at startup. +

+ If the Erlang runtime system is the only operating system + process that binds threads to logical processors, this + improves the performance of the runtime system. However, + if other operating system processes (as for example + another Erlang runtime system) also bind threads to + logical processors, there might be a performance penalty + instead. If this is the case you, are are advised to + unbind the schedulers using the + +sbtu command line argument, + or by invoking + erlang:system_flag(scheduler_bind_type, + unbound).

+ @@ -657,6 +677,22 @@ that the +sct flag may have to be passed before the +sbt flag on the command line (in case no CPU topology has been automatically detected).

+

The runtime system will by default bind schedulers to logical + processors using the default_bind bind type if the amount + of schedulers are at least equal to the amount of logical + processors configured, binding of schedulers is supported, + and a CPU topology is available at startup. +

+

NOTE: If the Erlang runtime system is the only operating + system process that binds threads to logical processors, this + improves the performance of the runtime system. However, if other + operating system processes (as for example another Erlang runtime + system) also bind threads to logical processors, there might be a + performance penalty instead. If this is the case you, are advised + to unbind the schedulers using the +sbtu command line + argument, or by invoking + erlang:system_flag(scheduler_bind_type, + unbound).

For more information, see erlang:system_flag(scheduler_bind_type, SchedulerBindType).

diff --git a/erts/doc/src/erlang.xml b/erts/doc/src/erlang.xml index c3c0dd77d7..c8fe7b4bd2 100644 --- a/erts/doc/src/erlang.xml +++ b/erts/doc/src/erlang.xml @@ -5156,10 +5156,21 @@ true For more information on how to define the CPU topology, see erlang:system_flag(cpu_topology, CpuTopology).

-

NOTE: If other programs on the system have bound - to processors, e.g. another Erlang runtime system, you - may lose performance when binding schedulers. Therefore, - schedulers are by default not bound.

+

The runtime system will by default bind schedulers to logical + processors using the default_bind bind type if the amount + of schedulers are at least equal to the amount of logical + processors configured, binding of schedulers is supported, + and a CPU topology is available at startup. +

+

NOTE: If the Erlang runtime system is the only + operating system process that binds threads to logical processors, + this improves the performance of the runtime system. However, + if other operating system processes (as for example another Erlang + runtime system) also bind threads to logical processors, there + might be a performance penalty instead. If this is the case you, + are are advised to unbind the schedulers using the + +sbtu command line argument, + or erlang:system_flag(scheduler_bind_type, unbound).

Schedulers can be bound in different ways. The How argument determines how schedulers are bound. How can currently be one of:

diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index e329bfd2d1..c61fab5d36 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -180,6 +180,7 @@ static ErtsCpuBindData *scheduler2cpu_map; erts_smp_rwmtx_t erts_cpu_bind_rwmtx; typedef enum { + ERTS_CPU_BIND_UNDEFINED, ERTS_CPU_BIND_SPREAD, ERTS_CPU_BIND_PROCESSOR_SPREAD, ERTS_CPU_BIND_THREAD_SPREAD, @@ -190,6 +191,9 @@ typedef enum { ERTS_CPU_BIND_NONE } ErtsCpuBindOrder; +#define ERTS_CPU_BIND_DEFAULT_BIND \ + ERTS_CPU_BIND_THREAD_NO_NODE_PROCESSOR_SPREAD + ErtsCpuBindOrder cpu_bind_order; static erts_cpu_topology_t *user_cpudata; @@ -3503,14 +3507,15 @@ erts_init_scheduler_bind_type(char *how) if (!system_cpudata && !user_cpudata) return ERTS_INIT_SCHED_BIND_TYPE_ERROR_NO_CPU_TOPOLOGY; - if (sys_strcmp(how, "s") == 0) + if (sys_strcmp(how, "db") == 0) + cpu_bind_order = ERTS_CPU_BIND_DEFAULT_BIND; + else if (sys_strcmp(how, "s") == 0) cpu_bind_order = ERTS_CPU_BIND_SPREAD; else if (sys_strcmp(how, "ps") == 0) cpu_bind_order = ERTS_CPU_BIND_PROCESSOR_SPREAD; else if (sys_strcmp(how, "ts") == 0) cpu_bind_order = ERTS_CPU_BIND_THREAD_SPREAD; - else if (sys_strcmp(how, "db") == 0 - || sys_strcmp(how, "tnnps") == 0) + else if (sys_strcmp(how, "tnnps") == 0) cpu_bind_order = ERTS_CPU_BIND_THREAD_NO_NODE_PROCESSOR_SPREAD; else if (sys_strcmp(how, "nnps") == 0) cpu_bind_order = ERTS_CPU_BIND_NO_NODE_PROCESSOR_SPREAD; @@ -4153,14 +4158,15 @@ erts_bind_schedulers(Process *c_p, Eterm how) old_cpu_bind_order = cpu_bind_order; - if (ERTS_IS_ATOM_STR("spread", how)) + if (ERTS_IS_ATOM_STR("default_bind", how)) + cpu_bind_order = ERTS_CPU_BIND_DEFAULT_BIND; + else if (ERTS_IS_ATOM_STR("spread", how)) cpu_bind_order = ERTS_CPU_BIND_SPREAD; else if (ERTS_IS_ATOM_STR("processor_spread", how)) cpu_bind_order = ERTS_CPU_BIND_PROCESSOR_SPREAD; else if (ERTS_IS_ATOM_STR("thread_spread", how)) cpu_bind_order = ERTS_CPU_BIND_THREAD_SPREAD; - else if (ERTS_IS_ATOM_STR("default_bind", how) - || ERTS_IS_ATOM_STR("thread_no_node_processor_spread", how)) + else if (ERTS_IS_ATOM_STR("thread_no_node_processor_spread", how)) cpu_bind_order = ERTS_CPU_BIND_THREAD_NO_NODE_PROCESSOR_SPREAD; else if (ERTS_IS_ATOM_STR("no_node_processor_spread", how)) cpu_bind_order = ERTS_CPU_BIND_NO_NODE_PROCESSOR_SPREAD; @@ -4206,14 +4212,15 @@ erts_fake_scheduler_bindings(Process *p, Eterm how) int cpudata_size; Eterm res; - if (ERTS_IS_ATOM_STR("spread", how)) + if (ERTS_IS_ATOM_STR("default_bind", how)) + fake_cpu_bind_order = ERTS_CPU_BIND_DEFAULT_BIND; + else if (ERTS_IS_ATOM_STR("spread", how)) fake_cpu_bind_order = ERTS_CPU_BIND_SPREAD; else if (ERTS_IS_ATOM_STR("processor_spread", how)) fake_cpu_bind_order = ERTS_CPU_BIND_PROCESSOR_SPREAD; else if (ERTS_IS_ATOM_STR("thread_spread", how)) fake_cpu_bind_order = ERTS_CPU_BIND_THREAD_SPREAD; - else if (ERTS_IS_ATOM_STR("default_bind", how) - || ERTS_IS_ATOM_STR("thread_no_node_processor_spread", how)) + else if (ERTS_IS_ATOM_STR("thread_no_node_processor_spread", how)) fake_cpu_bind_order = ERTS_CPU_BIND_THREAD_NO_NODE_PROCESSOR_SPREAD; else if (ERTS_IS_ATOM_STR("no_node_processor_spread", how)) fake_cpu_bind_order = ERTS_CPU_BIND_NO_NODE_PROCESSOR_SPREAD; @@ -4438,7 +4445,7 @@ early_cpu_bind_init(void) (sizeof(erts_cpu_topology_t) * system_cpudata_size)); - cpu_bind_order = ERTS_CPU_BIND_NONE; + cpu_bind_order = ERTS_CPU_BIND_UNDEFINED; if (!erts_get_cpu_topology(erts_cpuinfo, system_cpudata) || ERTS_INIT_CPU_TOPOLOGY_OK != verify_topology(system_cpudata, @@ -4464,6 +4471,17 @@ late_cpu_bind_init(void) scheduler2cpu_map[ix].bound_id = -1; } + if (cpu_bind_order == ERTS_CPU_BIND_UNDEFINED) { + int ncpus = erts_get_cpu_configured(erts_cpuinfo); + if (ncpus < 1 || erts_no_schedulers < ncpus) + cpu_bind_order = ERTS_CPU_BIND_NONE; + else + cpu_bind_order = ((system_cpudata || user_cpudata) + && (erts_bind_to_cpu(erts_cpuinfo, -1) != -ENOTSUP) + ? ERTS_CPU_BIND_DEFAULT_BIND + : ERTS_CPU_BIND_NONE); + } + if (cpu_bind_order != ERTS_CPU_BIND_NONE) { erts_cpu_topology_t *cpudata; int cpudata_size; -- cgit v1.2.3