Re: [PATCH 0/3] timer: Improve itimers scalability

[Hideaki Kimura <hideaki.kimura@hpe.com>]

On 08/26/2015 04:13 PM, Frederic Weisbecker wrote:
> On Wed, Aug 26, 2015 at 03:53:26PM -0700, Hideaki Kimura wrote:
>> Sure, let me elaborate.
>>
>> Executive summary:
>>   Yes, enabling a process-wide timer in such a large machine is not wise, but
>> sometimes users/applications cannot avoid it.
>>
>>
>> The issue was observed actually not in a database itself but in a common
>> library it links to; gperftools.
>>
>> The database itself is optimized for many-cores/sockets, so surely it avoids
>> putting a process-wide timer or other unscalable things. It just links to
>> libprofiler for an optional feature to profile performance bottleneck only
>> when the user turns it on. We of course avoid turning the feature on unless
>> while we debug/tune the database.
>>
>> However, libprofiler sets the timer even when the client program doesn't
>> invoke any of its functions: libprofiler does it when the shared library is
>> loaded. We requested the developer of libprofiler to change the behavior,
>> but seems like there is a reason to keep that behavior:
>>    https://code.google.com/p/gperftools/issues/detail?id=133
>>
>> Based on this, I think there are two reasons why we should ameliorate this
>> issue in kernel layer.
>>
>>
>> 1. In the particular case, it's hard to prevent or even detect the issue in
>> user space.
>>
>> We (a team of low-level database and kernel experts) in fact spent huge
>> amount of time to just figure out what's the bottleneck there because
>> nothing measurable happens in user space. I pulled out countless hairs.
>>
>> Also, the user has to de-link the library from the application to prevent
>> the itimer installation. Imagine a case where the software is proprietary.
>> It won't fly.
>>
>>
>> 2. This is just one example. There could be many other such
>> binaries/libraries that do similar things somewhere in a complex software
>> stack.
>>
>> Today we haven't heard of many such cases, but people will start hitting it
>> once 100s~1,000s of cores become common.
>>
>>
>> After applying this patchset, we have observed that the performance hit
>> almost completely went away at least for 240 cores. So, it's quite
>> beneficial in real world.
>
> I can easily imagine that many code incidentally use posix cpu timers when
> it's not strictly required. But it doesn't look right to fix the kernel
> for that. For this simple reason: posix cpu timers, even after your fix,
> should introduce noticeable overhead. All threads of a process with a timer
> enqueued in elapse the cputime in a shared atomic variable. Add to that the
> overhead of enqueuing the timer, firing it. There is a bunch of scalability
> issue there.

I totally agree that this is not a perfect solution. If there are 10x 
more cores and sockets, just the atomic fetch_add might be too expensive.

However, it's comparatively/realistically the best thing we can do 
without any drawbacks. We can't magically force all library developers 
to write the most scalable code always.

My point is: this is a safety net, and a very effective one.

-- 
Hideaki Kimura