Sandybridge Scalability with 3 & 4 cores

[Matthew Tippett]

Thanks Kris,

Moving to freebsd-performance.

Hi all,

When running NASA's NAS Parallel Benchmarks [1] under Phoronix Test 
Suite [2] on PC-BSD 8.1, we are seeing a scalability issue on Intel's 
new Sandybridge hardware.

We see scalability at 1, 2 and 6 processors enabled (via BIOS).  At 3 & 
4 cores, we're seeing drop off in some tests, or a collapse in 
performance in others.  Although it may be a quirk on the Sandybridge 
hardware, it may also be a scalability issue in general.

Attached are two images showing the collapse in scalability.  The LU.A 
test was unable to complete in a reasonable time (hours, vs the expected 
minutes) and was aborted.  Interestingly the scalability at 6 cores were 
in line with the other platforms.

To repeat these tests, you can download Phoronix Test Suite 3 RC and run 
the NAS Parallel Benchmarks via the command

       ./phoronix-test-suite benchmark pts/npb

And choose either the MG.B or LU.A subtest.  We'll be trying to remove 
the HW from being a potential factor.

Also note that scalability of 6 cores + HyperThreading also dropped off 
fairly heavily (so did OpenIndiana).

Thanks in advance.

Matthew

[1] http://www.nas.nasa.gov/Resources/Software/npb.html
[2] http://www.phoronix-test-suite.com/

On 02/14/2011 06:16 AM, Kris Moore wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> On 02/14/2011 02:09, Matthew Tippett wrote:
>> Hi Kris,
>>
>> I work with Michael Larabel, of Phoronix fame.  We've recently run some
>> benchmarks with a core i7 970 against OpenIndiana,  PC-BSD and Ubuntu
>> with various cores turned off.
>>
>> One interesting behaviour that we saw is that between 3 and 4 cores
>> running the performance of PC-BSD dropped fairly badly in some of the
>> NAS parallel benchmarks.  The system performs reasonably well for 1,2
>> and 6 cores.   Is this performance behaviour known within the PC-BSD
>> community?  It is fairly easy to reproduce the results with the Phoronix
>> Test Suite and we'd be happy to work with some developers in determining
>> the underlying issue.
>>
>> Note, however that the article covering this benchmark will probably get
>> published over the next few days.
>>
>> Regards,
>>
>> Matthew
> Matthew,
>
> Thanks for the heads up! What specific types of benchmarks are falling
> behind? On the PC-BSD side we mostly handle usability stuff, and don't
> monkey with the FreeBSD internals, so this may best be a question for
> the FreeBSD Performance mailing list. I'm sure you'll get all kinds of
> comments there on how to improve the scalability tuning, and if there is
> still an issue then those will be the people who can fix it :)
>
> http://lists.freebsd.org/mailman/listinfo/freebsd-performance
>
>
>
> - --
> Kris Moore
> PC-BSD Software
> iXsystems
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v2.0.16 (FreeBSD)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>
> iQEcBAEBAgAGBQJNWTlWAAoJEDv6T4U6J2HSvOUH/0NBW+8J/6JbLl4RFfY75Ujs
> Y7fAkqK8jmMVdaFxnwDm/wt3ThmWOcvzyCg+f/XUDDm41D9Js+iFbQlY8r8vR+Q6
> MwubRsw4eFJgS7fNTomakBw+NgiJwG1pFGyoTjWTz5VR3f3lMGBpkeFR40kT90eb
> K/Rn7f52ph1LInKyYZ4SF7NuBoF4akBfr2AGx7xvwoxaxNn9tJ92u/2jkHe//7k9
> GbQJ1f77hXMTsFkoYB0F3E4JEQuUzgaZWqodff8+gYKWBs0lsnhzL82HbKdfN3LI
> Oy5WOv0d2wNuQyYaT6iQ0r5xAZy3LZX6LL2qhwnHKs/dbkLSDwy/EtwwDlSNZUg=
> =jj6I
> -----END PGP SIGNATURE-----