[Scons-users] visualize build parallelism, duration of build steps?

[Gabe Black]

That sounds interesting. I see that when I set -j to 24, the actual CPU
percentage it gets is only 1953%, which I think would correlate with
underused CPUs. I had other things running on this computer in the mean
time like my web browser, but not enough to use 14 of 24 threads in the
-j12 case, for instance.

jobs  CPU percentage
4     397.00%
8     755.00%
12    1086.00%
16    1403.00%
20    1713.00%
24    1953.00%

On Tue, Oct 19, 2021 at 5:21 PM Bill Deegan <bill at baddogconsulting.com>
wrote:

> From some experiments MongoDB ran, you can saturate the scheduler without
> maxing out available CPU's with the current taskmaster.
> They produced a PR with some patches to improve the job finding and
> dispatch logic.
>
> If you're hitting a wall where you CPU cores are idle, that PR's patch
> might be interesting.
> Though it could be that SCons schedules lots of object compiles -> fewer
> ar/shared library linking -> fewer programs could also be leading to under
> utilized cores later in your build and/or as part of an incremental build.
>
> Hope that helps,
> Bill
> SCons Project Co-Manager
>
> On Tue, Oct 19, 2021 at 3:55 PM Gabe Black <gabe.black at gmail.com> wrote:
>
>> There definitely seems to be a build size component to this. I have a
>> series of changes which don't change how anything is built, but tell SCons
>> about everything and then select what to build with different dependencies
>> instead of using python logic to selectively set up build rules. With -j
>> set to its previous saturation point performance wise, this actually
>> increased build time by about 7-8%. I deleted the build and reran to make
>> sure there weren't caching effects impacting the results, and they stayed
>> consistent.
>>
>> Gabe
>>
>> On Tue, Oct 19, 2021 at 6:59 AM Mats Wichmann <mats at wichmann.us> wrote:
>>
>>> On 10/18/21 16:45, Gabe Black wrote:
>>> > Hi folks, sorry if this is a really obvious question, but is there a
>>> > command line flag or tool or something to visualize how parallel a
>>> SCons
>>> > build is, if there are any bottlenecks, if there are abnormally long
>>> > running build steps, etc?
>>>
>>> The taskmastertrace is generally the way to extract this information.
>>> It's a bit like trying to drink from a firehose, though, so a tool to
>>> process it would likely help. Don't know if there are any of those
>>> floating around.
>>>
>>>
>>> > I've had a suspicion for a few years that there might be a task
>>> scheduling bug where at certain points a super task seems to come along and
>>> stop any other tasks from being scheduled until it is done.
>>> There's been some discussion of this over time.  It appears the way
>>> scheduling is done may be prone to getting blocked by a long-running
>>> jobs.  One tuning of the algorithm was proposed in a PR which hasn't
>>> received much action - opinions about whether this had really pinpointed
>>> an underlying issue were not unanimous.
>>>
>>> https://github.com/SCons/scons/pull/3386
>>>
>>> There's also been talk about maybe not having all the members of the
>>> pool of job threads be handled identically, so there's a place you could
>>> put a long-running job to get it started as soon as possible, and not
>>> wait behind a bunch of shorter ones, which could then compete for the
>>> other threads and maybe get closer to finishing at the same times.
>>> Making decisions based on length of jobs would require some sort of
>>> profiling capability so the scheduler has the information available.
>>>
>>> On the whole I think it's been posited that there are performance issues
>>> when the build gets very big, and not sure it's well understood why.
>>> The taskmaster is a pretty complex piece of code and it was heavily
>>> tuned "in the old days" (~15 years ago).  Be nice to have better ways to
>>> characterize what happens on modern hardware configurations with
>>> potentially larger build projects than when that work was done.
>>>
>>> _______________________________________________
>> Scons-users mailing list
>> Scons-users at scons.org
>> https://pairlist4.pair.net/mailman/listinfo/scons-users
>>
> _______________________________________________
> Scons-users mailing list
> Scons-users at scons.org
> https://pairlist4.pair.net/mailman/listinfo/scons-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://pairlist4.pair.net/pipermail/scons-users/attachments/20211019/b37c0ebd/attachment.htm>