Home > Erlang探索, 源码分析, 网络编程, 调优 > Erlang 网络密集型服务器的瓶颈和解决思路

Erlang 网络密集型服务器的瓶颈和解决思路

November 11th, 2013

原创文章,转载请注明: 转载自系统技术非业余研究

本文链接地址: Erlang 网络密集型服务器的瓶颈和解决思路

最近我们的Erlang IO密集型的服务器程序要做细致的性能提升,从每秒40万包处理提升到60万目标,需要对进程和IO调度器的原理很熟悉,并且对行为进行微调,花了不少时间参阅了相关的文档和代码。

1. Characterizing the Scalability of Erlang VM on Many-core Processors 参见这里
2. Evaluate the benefits of SMP support for IO-intensive Erlang applications 参见这里

我们的性能瓶颈目前根据 lcnt 的提示:

1. 调度器运行队列的锁冲突,参见下图:

2. erlang只有单个poll set, 大量的IO导致性能瓶颈,摘抄“Evaluate the benefits of SMP support for IO-intensive Erlang applications” P46的结论如下:

Finally, we analyzed how IO operations are handled by the Erlang VM, and that
was the bottleneck. The problem relies on the fact that there is only one global
poll-set where IO tasks are kept. Hence, only one scheduler at a time can call
erts_check_io (the responsible function for performing IO tasks) to obtain pending
tasks from the poll-set. So, a scheduler can finish its job, but it has to wait
idly for the other schedulers to complete their IO pending tasks before it starts
it own ones. In more details, for N scheduler, only one can call erts_check_io
regardless the load; the other N-1 schedulers will start spinning until they gain
access to erts_check_io() and finish executing their IO tasks. For a bigger number of
schedulers, more schedulers will spin, and more CPU time will be waisted on spinning.
This behavior was noticed even during our evaluations when running the tests in a 8
cores machine, apart from the 16 cores one. There are two conditions that determine
whether a scheduler can access erts_check_io, one is for a mutex variable named
“doing_sys_schedule” to be unlocked, and the other one is to make sure the variable
“erts_port_task_outstanding_io_tasks” reaches a value of 0 meaning that there is
no IO task to process in the whole system. If one of the conditions breaks and there
is no other task to process, the scheduler starts spinning. Emysql driver generates
a lot of processes to be executed by the schedulers since a new process is spawned
per each single insert or read request. By using Erlang etop tool, we can check the
lifetime of all the processes created in the Erlang VM. Their lifetime is extremely
short, and they do nothing else (CPU-related) apart from the IO requests. The
requests are serialized at this point because whenever a scheduler starts doing the
system scheduling, it locks the mutex variable “doing_sys_schedule”, and then calls
erts_check_io() function for finding pending IO tasks. These tasks are spread to the
other schedulers and are processed only when the two aforementioned conditions are

这之前,我们提了相关的patch到erlang/otp团队了,但是没有获准合并,讨论见这里。官方承诺的multiple poll set功能都好几年也没见实现。

其中IO调度器的迁移逻辑,见第一篇paper, 简单的描述可以见这里

They migration logic does:
collect statistics about the maxlength of all scheduler’s runqueues
setup migration paths
Take away jobs from full-load schedulers and pushing jobs on low load scheduler queues
Running on full load or not! If all schedulers are not fully loaded, jobs will be migrated to schedulers with lower id’s and thus making some schedulers inactive.


$ erl
Erlang R17A (erts-5.11) [source-e917f6d] [64-bit] [smp:16:16] [async-threads:10] [hipe] [kernel-poll:false] [type-assertions] [debug-compiled] [lock-checking] [systemtap]

Eshell V5.11  (abort with ^G)
1> erlang:system_flag(scheduling_statistics,enable).  
2> erlang:system_info(scheduling_statistics).         
3> erlang:system_info(total_scheduling_statistics).

对应下erl_process.c实现源码,我们可以知道数字的含义: 第一个数字调度器代表执行的进程数目,第二个为迁移的进程数目。

        for (i = 0; i < ERTS_NO_PRIO_LEVELS; i++) {
            prio[i] = erts_sched_stat.prio[i].name;
            executed[i] = erts_sched_stat.prio[i].total_executed;
            migrated[i] = erts_sched_stat.prio[i].total_migrated;

特别是迁移进程的时候,会引起大量的目标和源队列锁操作。如果观察到迁移的进程数目过多的话,我们就要考虑未公开的进程绑定选项{scheduler, N}在进程创建的时候把某个进程固定在调度器上,避免迁移。 具体操作可以参考这篇


Post Footer automatically generated by wp-posturl plugin for wordpress.

  1. November 11th, 2013 at 21:08 | #1


    Yu Feng Reply:


Comments are closed.