SMP & HTT on 6.3 (P4)

classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

SMP & HTT on 6.3 (P4)

Todor Todorov
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi,
what do you think for HyperThreading (P4 GHz), which serves FBSD 6.3?
Now it's disabled by the BIOS but since today I've upgraded the machine
5.5 to 6.3 though if under 6.XX series it worths or not.

I've read for performance penalties under 5.XX series w/ HTT on. I'm not
going to change to ULE (it's said to be better for HTT awareness).

Please advice.

If someone has a machine nearby please paste dmesg & top.

Regards,
Todor
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.4 (FreeBSD)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFHp5raibJkIG65HMcRAkw3AKCSAoL0DATESNTW5+Bm8n6hAlHrAwCgw4Y3
+NEJKcuQwQprYU7IgrK4ztc=
=Dd8d
-----END PGP SIGNATURE-----
_______________________________________________
[hidden email] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "[hidden email]"
Reply | Threaded
Open this post in threaded view
|

Re: SMP & HTT on 6.3 (P4)

Jeremie Le Hen-2
Hi,

On Tue, Feb 05, 2008 at 01:08:10AM +0200, Todorov wrote:

>  -----BEGIN PGP SIGNED MESSAGE-----
>  Hash: SHA1
>
>  Hi,
>  what do you think for HyperThreading (P4 GHz), which serves FBSD 6.3?
>  Now it's disabled by the BIOS but since today I've upgraded the machine
>  5.5 to 6.3 though if under 6.XX series it worths or not.
>
>  I've read for performance penalties under 5.XX series w/ HTT on. I'm not
>  going to change to ULE (it's said to be better for HTT awareness).

ULE is broken on RELENG_6.  Forget it.  If you really want ULE, you'll
have to upgrade to RELENG_7.

Regarding HTT, I haven't heard of any performance improvement with HTT
on FreeBSD.  The scheduler have to be aware of it because HTT processors
share the same cache lines (including L1).  Unenlightened scheduling
leads to an increase of cache miss (42% according [1]).  This is not the
case on any FreeBSD branch currently.  Nonetheless Jeff Roberson is
working on a CPU topology-aware scheduler implementation [2] but this is
not even in -CURRENT yet.  It will probably ends up in RELENG_8.

Besides, Colin Percival has shown that HTT could lead to local
information leak [3].

WRT to these informations, I advice you to disable HTT.

Regards,

[1] http://www.theinquirer.net/en/inquirer/news/2006/08/02/arm-is-no-fan-of-hyperthreading
[2] http://jeffr-tech.livejournal.com/17426.html
[3] http://www.daemonology.net/hyperthreading-considered-harmful/
--
Jeremie Le Hen
< jeremie at le-hen dot org >< ttz at chchile dot org >
_______________________________________________
[hidden email] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "[hidden email]"
Reply | Threaded
Open this post in threaded view
|

Re: SMP & HTT on 6.3 (P4)

Kris Kennaway-3
Jeremie Le Hen wrote:

> Hi,
>
> On Tue, Feb 05, 2008 at 01:08:10AM +0200, Todorov wrote:
>>  -----BEGIN PGP SIGNED MESSAGE-----
>>  Hash: SHA1
>>
>>  Hi,
>>  what do you think for HyperThreading (P4 GHz), which serves FBSD 6.3?
>>  Now it's disabled by the BIOS but since today I've upgraded the machine
>>  5.5 to 6.3 though if under 6.XX series it worths or not.
>>
>>  I've read for performance penalties under 5.XX series w/ HTT on. I'm not
>>  going to change to ULE (it's said to be better for HTT awareness).
>
> ULE is broken on RELENG_6.  Forget it.  If you really want ULE, you'll
> have to upgrade to RELENG_7.
>
> Regarding HTT, I haven't heard of any performance improvement with HTT
> on FreeBSD.  The scheduler have to be aware of it because HTT processors
> share the same cache lines (including L1).  Unenlightened scheduling
> leads to an increase of cache miss (42% according [1]).  This is not the
> case on any FreeBSD branch currently.  Nonetheless Jeff Roberson is
> working on a CPU topology-aware scheduler implementation [2] but this is
> not even in -CURRENT yet.  It will probably ends up in RELENG_8.

Actually with ULE in 7.0 it often seems to help on the workloads I have
tested.  This is probably because ULE has much better CPU affinity so
processes don't bounce around between CPU threads so much.

The bottom line with HTT has really always been: try it on your workload
and measure whether or not it helps.

Kris

_______________________________________________
[hidden email] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "[hidden email]"