Re: Futex

[Bhavin Thaker]

Yes, here are some links that confirm Haitao Wang’s statement:

The classic book, The linux programming interface, by Michael Kerrisk:
On Linux, mutexes are implemented using futexes(an acronym derived from fast
user space mutexes), and lock contentions are dealt with using the
futex()system
call.

https://stackoverflow.com/questions/6364314/why-is-a-pthread-mutex-considered-slower-than-a-futex

Bhavin Thaker.


On Thu, Nov 23, 2017 at 10:00 PM kellen sunderland <
kellen.sunderl...@gmail.com> wrote:

> I think Haitao is right given some stacks we've recently looked at.  e.g.
> https://gist.github.com/KellenSunderland/893d11165e19d1efcf5c0fe8e8584600
>
> -Kellen
>
> On Thu, Nov 23, 2017 at 7:32 PM, Haitao Wang <hai...@openailab.com> wrote:
>
> > Hi, Chris,
> >
> > As far as I know, the mutex implementation in Linux is based on futex
> > already.
> >
> > Thanks,
> > Haitao
> >
> > -Original Message-
> > From: dev-return-1553-haitao=openailab@mxnet.incubator.apache.org
> > [mailto:dev-return-1553-haitao=openailab@mxnet.incubator.apache.org]
> > On Behalf Of Chris Olivier
> > Sent: Friday, November 24, 2017 3:02 AM
> > To: dev@mxnet.incubator.apache.org
> > Subject: Futex
> >
> > Was doing some timing with futexes (we used them a lot in a previous life
> > in database engines) and they're consistently about 20-30% faster than
> > standard mutexes in Linux.
> >
> > However, it seems like this is not worth making a change since mutexes
> > don't tend to get called so much that it would seem to make a noticeable
> > difference, although I could be wrong -- so far besides the queue, I am
> not
> > aware of any major bottlenecks on mutexes.
> >
> > Any thoughts?
> >
> > -Chris
> >
> >
>

Re: Futex

[kellen sunderland]

I think Haitao is right given some stacks we've recently looked at.  e.g.
https://gist.github.com/KellenSunderland/893d11165e19d1efcf5c0fe8e8584600

-Kellen

On Thu, Nov 23, 2017 at 7:32 PM, Haitao Wang <hai...@openailab.com> wrote:

> Hi, Chris,
>
> As far as I know, the mutex implementation in Linux is based on futex
> already.
>
> Thanks,
> Haitao
>
> -Original Message-
> From: dev-return-1553-haitao=openailab@mxnet.incubator.apache.org
> [mailto:dev-return-1553-haitao=openailab@mxnet.incubator.apache.org]
> On Behalf Of Chris Olivier
> Sent: Friday, November 24, 2017 3:02 AM
> To: dev@mxnet.incubator.apache.org
> Subject: Futex
>
> Was doing some timing with futexes (we used them a lot in a previous life
> in database engines) and they're consistently about 20-30% faster than
> standard mutexes in Linux.
>
> However, it seems like this is not worth making a change since mutexes
> don't tend to get called so much that it would seem to make a noticeable
> difference, although I could be wrong -- so far besides the queue, I am not
> aware of any major bottlenecks on mutexes.
>
> Any thoughts?
>
> -Chris
>
>

RE: Futex

[Haitao Wang]

Hi, Chris,

As far as I know, the mutex implementation in Linux is based on futex already.

Thanks,
Haitao

-Original Message-
From: dev-return-1553-haitao=openailab@mxnet.incubator.apache.org 
[mailto:dev-return-1553-haitao=openailab@mxnet.incubator.apache.org] On 
Behalf Of Chris Olivier
Sent: Friday, November 24, 2017 3:02 AM
To: dev@mxnet.incubator.apache.org
Subject: Futex

Was doing some timing with futexes (we used them a lot in a previous life in 
database engines) and they're consistently about 20-30% faster than standard 
mutexes in Linux.

However, it seems like this is not worth making a change since mutexes don't 
tend to get called so much that it would seem to make a noticeable difference, 
although I could be wrong -- so far besides the queue, I am not aware of any 
major bottlenecks on mutexes.

Any thoughts?

-Chris

Re: Futex

[Chris Olivier]

Is there a claim of a mutex bottleneck in mxnet? I am not aware of any,
unless someone knows otherwise.

Database engines tend to be collision-happy, but mxnet does not seem to be
as far as I know, unless someone knows differently.


On Thu, Nov 23, 2017 at 11:06 AM, Bhavin Thaker 
wrote:

> Do we have performance/profiling measurement evidence to prove/disprove the
> claims of the impact of mutex bottleneck in MXNet?
>
> Bhavin Thaker.
>
> On Thu, Nov 23, 2017 at 11:03 AM Tianqi Chen 
> wrote:
>
> > If they are not bottleneck. Then resorting to standard library solution
> is
> > always preferred
> >
> > Tianqi
> >
> > On Thu, Nov 23, 2017 at 11:01 AM, Chris Olivier 
> > wrote:
> >
> > > Was doing some timing with futexes (we used them a lot in a previous
> life
> > > in database engines) and they're consistently about 20-30% faster than
> > > standard mutexes in Linux.
> > >
> > > However, it seems like this is not worth making a change since mutexes
> > > don't tend to get called so much that it would seem to make a
> noticeable
> > > difference, although I could be wrong -- so far besides the queue, I am
> > not
> > > aware of any major bottlenecks on mutexes.
> > >
> > > Any thoughts?
> > >
> > > -Chris
> > >
> >
>

Re: Futex

[Bhavin Thaker]

Do we have performance/profiling measurement evidence to prove/disprove the
claims of the impact of mutex bottleneck in MXNet?

Bhavin Thaker.

On Thu, Nov 23, 2017 at 11:03 AM Tianqi Chen 
wrote:

> If they are not bottleneck. Then resorting to standard library solution is
> always preferred
>
> Tianqi
>
> On Thu, Nov 23, 2017 at 11:01 AM, Chris Olivier 
> wrote:
>
> > Was doing some timing with futexes (we used them a lot in a previous life
> > in database engines) and they're consistently about 20-30% faster than
> > standard mutexes in Linux.
> >
> > However, it seems like this is not worth making a change since mutexes
> > don't tend to get called so much that it would seem to make a noticeable
> > difference, although I could be wrong -- so far besides the queue, I am
> not
> > aware of any major bottlenecks on mutexes.
> >
> > Any thoughts?
> >
> > -Chris
> >
>

Re: Futex

[Tianqi Chen]

If they are not bottleneck. Then resorting to standard library solution is
always preferred

Tianqi

On Thu, Nov 23, 2017 at 11:01 AM, Chris Olivier 
wrote:

> Was doing some timing with futexes (we used them a lot in a previous life
> in database engines) and they're consistently about 20-30% faster than
> standard mutexes in Linux.
>
> However, it seems like this is not worth making a change since mutexes
> don't tend to get called so much that it would seem to make a noticeable
> difference, although I could be wrong -- so far besides the queue, I am not
> aware of any major bottlenecks on mutexes.
>
> Any thoughts?
>
> -Chris
>

Futex

[Chris Olivier]

Was doing some timing with futexes (we used them a lot in a previous life
in database engines) and they're consistently about 20-30% faster than
standard mutexes in Linux.

However, it seems like this is not worth making a change since mutexes
don't tend to get called so much that it would seem to make a noticeable
difference, although I could be wrong -- so far besides the queue, I am not
aware of any major bottlenecks on mutexes.

Any thoughts?

-Chris