[erlang-questions] About Erlang SMP scheduler

Rickard Green <>
Fri Jun 1 10:05:45 CEST 2012


On 05/31/2012 01:29 PM, Emilio De Camargo Francesquini wrote:
> Hello,
> 
>> check_balance() sets up
>> migration paths and migration limits that are used in order to balance
>> the load between schedulers. Without balancing, processes with the
>> same assigned priority can get very different effective priorities.
> 
> I have some questions relating to that.
> 
> First, after the decision to rebalance has been made, what are the
> criteria to select which processes from each queue will be migrated?
> The last ones in the queue, the first, at random, ...?
> 

Each priority is balanced separately (while considering the effects
different priorities have on each other). Run-queue lengths on a
specific priority is the most important factor.  Within each priority
all processes are considered equal.

We migrate away processes from an overloaded run queue at (sort of) both
ends. New processes will be redirected instead of enqueued and
underloaded schedulers will pick from the front of the queue.

> Second, how are these migration paths and limits defined? Does the
> definition of the migration paths take into consideration the
> architecture (SMP, multisocket SMP, NUMA, ...) of the machine? If we
> consider that the lifespan of the majority of the process is quite
> short, and if we are running on a NUMA machine for example, migrating
> processes at random to another numa-node might hurt the performance
> more than letting the execution queues unbalanced.
> 

The primary purpose of the load balancing is to preserve the soft real
time properties. That is, a load balancing decision that hurt
performance isn't necessarily unwanted. We do, of course, not want to
unnecessarily hurt performance. A performance improvement is, of course,
preferred.

The CPU topology is currently taken into account in a situation where
schedulers run out of work. In this case we try to move work towards
schedulers with low ids. Scheduler bind type will determine the effect
of this compaction of work (for more info see cpu topology
<http://www.erlang.org/doc/man/erl.html#+sct> and scheduler bind type
<http://www.erlang.org/doc/man/erl.html#+sbt> in the documentation erl).

We have on the todo-list to take the CPU topology into account also in
other load situations. Most likely some kind of hierarchical balancing
strategy based on the CPU topology. However, at the time being there are
other things ahead of this on the todo-list, but at some point this will
be introduced.

Regards,
Rickard

> Thanks!
> 
> Best Regards
> 
> Emilio Francesquini
> 
> 2012/5/8 Rickard Green<>:
>> The work stealing is there to quickly distribute work between
>> schedulers, but it doesn't balance the load. check_balance() sets up
>> migration paths and migration limits that are used in order to balance
>> the load between schedulers. Without balancing, processes with the
>> same assigned priority can get very different effective priorities.
>>
>> Example: With 4 schedulers and 400 cpu bound processes executing the
>> same code, each scheduler will eventually end up with 100 processes to
>> manage. That is, all processes will have the same access to the cpu.
>> If you disable check_balance(), you may end up with three schedulers
>> only having one process each, and one scheduler with 397 processes.
>> These 397 processes will effectively have a much lower priority than
>> the other three processes.
>>
>> Regards,
>> Rickard Green, Erlang/OTP, Ericsson AB
>>
>> 2012/5/2 Siyao Zheng(郑思遥)<>:
>>> Hi,
>>>
>>> Yes, that's true. But the "try_steal_task()" afterwards is also controlled
>>> by ERTS_SMP macro. It doesn't make any sense to talk about workload balance
>>> outside multi-processor environment. I'm just wondering why "check_balance"
>>> is needed while there exists work stealing(try_steal_task).
>>>
>>> Cheers
>>>
>>> On May 2, 2012, at 12:04 PM, xu yu wrote:
>>>
>>> Hi,
>>>
>>> "check_balance" controlled by "ifdef ERTS_SMP", so...
>>>
>>>
>>> 2012/4/27 "Siyao Zheng(郑思遥)"<>
>>>>
>>>> Hi,
>>>>
>>>> schedule(), in which check_balance() is only called.
>>>>
>>>> In schedule(), when check_balance_reds reaches zero, scheduler decides to
>>>> "check_balance". But later in schedule(), when scheduler finds run queue
>>>> being empty, it will try to steal task from other schedulers (by calling
>>>> try_steal_task()), which is also a load balance mechanism. I'm just
>>>> wondering what the purpose of check_balance().
>>>>
>>>> Cheers
>>>> Siyao
>>>>
>>>> On Apr 27, 2012, at 9:40 PM, Zabrane Mickael wrote:
>>>>
>>>>> Hi Siyao,
>>>>>
>>>>> Which "check_balance" you've commented out?
>>>>>
>>>>> Here is what I found so far:
>>>>>
>>>>> $ find /otp_src_R15B01  -type f | xargs grep check_balance | grep -v
>>>>> matches
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:    int
>>>>> forced_check_balance;
>>>>>
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:check_balance(ErtsRunQueue
>>>>> *c_rq)
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:#  error
>>>>> check_balance() assumes ERTS_MAX_PROCESS<  (1<<  27)
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> c_rq->check_balance_reds = INT_MAX;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> c_rq->check_balance_reds = INT_MAX;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> rq->check_balance_reds = ERTS_RUNQ_CALL_CHECK_BALANCE_REDS;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:     *
>>>>> check_balance() is never called in more threads
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:    forced =
>>>>> balance_info.forced_check_balance;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>>   balance_info.forced_check_balance = 0;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> c_rq->check_balance_reds = INT_MAX;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> rq->check_balance_reds = INT_MAX;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> rq->check_balance_reds = ERTS_RUNQ_CALL_CHECK_BALANCE_REDS;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> rq->check_balance_reds = ERTS_RUNQ_CALL_CHECK_BALANCE_REDS;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>>   balance_info.forced_check_balance = 0;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>>   (RQ)->check_balance_reds = ERTS_RUNQ_CALL_CHECK_BALANCE_REDS;        \
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> balance_info.forced_check_balance = 1;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> ERTS_RUNQ_IX(0)->check_balance_reds = 0;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c: if
>>>>> (rq->check_balance_reds<= 0)
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.c:
>>>>> check_balance(rq);
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.h:    int
>>>>> check_balance_reds;
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.h:
>>>>>   (RQ)->check_balance_reds -= (REDS);                          \
>>>>> /opt/otp_src_R15B01/erts/emulator/beam/erl_process.h:
>>>>>   (RQ)->check_balance_reds -= (REDS);
>>>>> [...]
>>>>>
>>>>> For OTP Team: is this dangerous ?
>>>>>
>>>>> Regards,
>>>>> Zabrane
>>>>>
>>>>>
>>>>> On Apr 27, 2012, at 1:23 PM, Siyao Zheng(郑思遥) wrote:
>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> In SMP version of Erlang, every scheduler periodically call
>>>>>> "check_balance()" to check load balance among all schedulers, then it might
>>>>>> migrate some processes to balance load, or shut down some schedulers with
>>>>>> low load. Does anyone know why scheduler should do this? check_balance() is
>>>>>> quite a big one, and it has to lock every run queue when it checks the run
>>>>>> queue. I think it's quite a big cost here. The work stealing at each
>>>>>> schedule step afterwards actually does work load balance very well. Actually
>>>>>> after I comment out the check_balance() step and run BigBang and Hackbench
>>>>>> benchmarks, they are really faster!
>>>>>>
>>>>>> So, I wonder what is the purpose of check_balance() step here, if it is
>>>>>> related to scalability on many core processors or something, and if there is
>>>>>> any other benchmark I can run to prove its usefulness.
>>>>>>
>>>>>> Cheers!
>>>>>>
>>>>>> Siyao
>>>>>> _______________________________________________
>>>>>> erlang-questions mailing list
>>>>>> 
>>>>>> http://erlang.org/mailman/listinfo/erlang-questions
>>>>>
>>>>>
>>>>
>>>> _______________________________________________
>>>> erlang-questions mailing list
>>>> 
>>>> http://erlang.org/mailman/listinfo/erlang-questions
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> erlang-questions mailing list
>>> 
>>> http://erlang.org/mailman/listinfo/erlang-questions
>>>
>> _______________________________________________
>> erlang-questions mailing list
>> 
>> http://erlang.org/mailman/listinfo/erlang-questions


-- 
Rickard Green, Erlang/OTP, Ericsson AB.



More information about the erlang-questions mailing list