public inbox for cygwin@cygwin.com
 help / color / mirror / Atom feed
From: Brian Inglis <Brian.Inglis@Shaw.ca>
To: cygwin@cygwin.com
Cc: jeff <jeff@jeffunit.com>
Subject: Re: posix thread scaling issue
Date: Sat, 2 Sep 2023 13:59:01 -0600	[thread overview]
Message-ID: <07386659-68b3-a35d-1402-22684f8e5755@Shaw.ca> (raw)
In-Reply-To: <cf618819-c30c-439d-ad5f-54b2311bd936@jeffunit.com>

On 2023-09-02 12:27, jeff via Cygwin wrote:
> On 9/2/2023 10:56, Brian Inglis wrote:
>> On 2023-09-02 08:57, jeff via Cygwin wrote:
>>> I have a program that is embarrassing parallel.
>>> On my older computer which has an epyc 7302 (16 cores,  32 threads) it scales 
>>> very well using cygwin, and fully utilized all threads.
>>> On my new computer which has an epyc 7B13 (64 cores, 128 threads) it does not 
>>> scale very well.
>>> According to the windows task manager, it only uses 74% of the cpu resources.
>>> The time it takes the program to run on windows is 166 seconds.
>>> Using the same hardware on a recent version of linux, I can get 100% cpu 
>>> utilization and the program takes 100 seconds to run.
>>> I suspect there may be something in cygwin that doesn't scale well with lots 
>>> of posix threads.

Both Windows and Cygwin support multiple processor groups, as some developers, 
maintainers, and users need support on such systems, and the process and thread 
support has been added to Cygwin.

>>> I know this is a bit of an unusual situation, but you can buy a 128 core / 
>>> 256 thread system now.
>>> Enclosed is the output of cygcheck.
>>> I updated my version of cygwin to be current as of today, Sep 2 2023.

>> What Windows edition and version are you running?
>> For details run:
>>
>> $ reg query "HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows NT\CurrentVersion" \
>> | sed '/^\s\+\.*\s/!d;/^.\{80,\}/d'
>>
>> Some retail editions limit you to 64 threads and that seems to be your case:
>>
>>     NUMBER_OF_PROCESSORS = '64'
>>
>> To make full use of your processors, you may have to upgrade your Windows to a 
>> commercial licence (and installation) of Windows 10/11 Pro for Workstations, 
>> enabling server features on non-server "Worskations" ~ HEDTs (High-End 
>> DeskTops); see:
>>
>> https://www.anandtech.com/show/15483/amd-threadripper-3990x-review/3
>>
>> or just run Linux!
>>
>> Watch out for terms misused like processor == socket on some sites!
>>
>> Also, you have to consider these are server systems, mainly designed for VM 
>> not HPC (High Performance Computing) parallelism.
>>
>> Your older system has higher base and boost/turbo clocks 3.0-3.3GHz: your 
>> newer system has lower clocks 2.25-2.65/3/3.5GHz which seems to depend on OEM 
>> target.
>>
>> You may also need to upgrade your memory, as each core could run ~10GB/s 
>> instructions, and these workstations are often provisioned with 128-256GB 
>> (2-4GB/core), so that may also need a Windows edition upgrade.

> I am running windows 10 professional. Using the task manager, 64 cores and 128 
> threads shows up for my processor.

As the linked AnandTech article shows and explains with Task Manager/ 
Performance tab, Win 10 Pro may think you have dual sockets, that limits the 
maximum thread parallelism you can achieve:

"Now the thing is, Workstation and Enterprise are built with multiple processor 
groups in mind, whereas Pro is not."

> Here is the output of your regex:
>      SystemRoot    REG_SZ    C:\Windows
>      BaseBuildRevisionNumber    REG_DWORD    0x1
>      BuildBranch    REG_SZ    vb_release
>      BuildGUID    REG_SZ    ffffffff-ffff-ffff-ffff-ffffffffffff
>      BuildLab    REG_SZ    19041.vb_release.191206-1406
>      BuildLabEx    REG_SZ    19041.1.amd64fre.vb_release.191206-1406
>      CompositionEditionID    REG_SZ    Enterprise
>      CurrentBuild    REG_SZ    19045
>      CurrentBuildNumber    REG_SZ    19045
>      CurrentMajorVersionNumber    REG_DWORD    0xa
>      CurrentMinorVersionNumber    REG_DWORD    0x0
>      CurrentType    REG_SZ    Multiprocessor Free
>      CurrentVersion    REG_SZ    6.3
>      EditionID    REG_SZ    Professional
>      EditionSubManufacturer    REG_SZ
>      EditionSubstring    REG_SZ
>      EditionSubVersion    REG_SZ
>      InstallationType    REG_SZ    Client
>      InstallDate    REG_DWORD    0x61e2300a
>      ProductName    REG_SZ    Windows 10 Pro
>      ReleaseId    REG_SZ    2009
>      SoftwareType    REG_SZ    System
>      UBR    REG_DWORD    0xcfc
>      PathName    REG_SZ    C:\Windows
>      ProductId    REG_SZ    00330-80000-00000-AA073
>      DisplayVersion    REG_SZ    22H2
>      RegisteredOwner    REG_SZ    jdeifik
>      RegisteredOrganization    REG_SZ
>      InstallTime    REG_QWORD    0x1d809b6d4ce7b09
> 
> In practice, but the new and old processors typically run at about 3ghz when 
> under load.
> When idling, both processors use about the same amount of power.
> 
> I have 128gb of ram, in 4 slots. Using that configuration, I can get 100% load 
> and significant faster performance on linux.
> Therefore I conclude the issue is either with windows or cygwin, and is not a 
> hardware issue.
 >
> When I run cinebench, I can get to 100% cpu utulization (at around 3ghz) on 
> windows.

Chances are the benchmark is designed to handle that:

"When the program is running inside the group, unless it is processor group 
aware, then it can only access other threads in the same group. This means that 
if a multi-threaded program can use 128 threads, if it isn’t built with 
processor groups in mind, then it might only spawn with access to 64."

I also do not know how you would program for that in Cygwin to map onto the 
equivalent Windows function required.

Perhaps one of the developers involved could comment here?

> As for what the processors are 'designed' for, I really don't care.
> I want a reliable, fast computer with ECC memory, and I can get that with an 
> EPYC processor.
> If a workload needs more than 128gb of memory, you pretty much need to use 
> server processors.
> I can put in up to 2tb of memory in my system, if I have the need for that.

As I suggested above, and as the AT tests suggest, with your configuration, you 
may get better results disabling multithreading on your current system, or 
running Pro for Workstations, which you may be able to test using a generic key.

Pro for Workstations is used and recommended by video shops, with much lower 
costs and power consumption running AMD than Intel, as a designer's workstation 
alternative getting better performance and reponsiveness than using servers for 
the same task.

-- 
Take care. Thanks, Brian Inglis              Calgary, Alberta, Canada

La perfection est atteinte                   Perfection is achieved
non pas lorsqu'il n'y a plus rien à ajouter  not when there is no more to add
mais lorsqu'il n'y a plus rien à retirer     but when there is no more to cut
                                 -- Antoine de Saint-Exupéry

  reply	other threads:[~2023-09-02 19:59 UTC|newest]

Thread overview: 10+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-09-02 14:57 jeff
2023-09-02 17:56 ` Brian Inglis
2023-09-02 18:27   ` jeff
2023-09-02 19:59     ` Brian Inglis [this message]
2023-09-02 20:04       ` jeff
2023-09-03  6:13         ` ASSI
2023-09-03  3:50       ` Mark Geisert
2023-09-03  4:13         ` Mark Geisert
2023-09-02 19:30 André Bleau
     [not found] ` <e36d50d5-75d0-40d5-92e2-02d04092fd77@jeffunit.com>
2023-09-02 21:23   ` André Bleau

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=07386659-68b3-a35d-1402-22684f8e5755@Shaw.ca \
    --to=brian.inglis@shaw.ca \
    --cc=cygwin@cygwin.com \
    --cc=jeff@jeffunit.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).