Skip site navigation (1)Skip section navigation (2)
Date:      24 Oct 97 10:13:27 -0700
From:      "Tina Yang" <TYANG@nc.com>
To:        shieyuan@eecs.harvard.edu, owner-freebsd-smp@freebsd.org
Cc:        freebsd-smp@freebsd.org
Subject:   Re: SMP is slower when two processors are enabled
Message-ID:  <199710241714.KAA12913@deimos.nc.com>

next in thread | raw e-mail | index | archive | help

--=_ORCL_2577328_0_1191971024111509_0
Content-Transfer-Encoding:quoted-printable
Content-Type:text/plain; charset="iso-8859-1"


>ShieYuan Wang said:
>> protocol stack to transfer unlimited data on the same local host. If
I
>> only enable one processor, the throughput can be 35 MB/sec. However
if
>> I enable the other processor hoping to get better throughput, I only
get
>> 12 MB/sec.
>>
>> I know that locking overhead between multiple processors may be a
lot. But
>> I never thought that it could be so high. Could anyone give me some
clues 
>> ABOUT THIS BAD Performance? My SMP 3.0 was installed about 6 months
ago.
>> Does the more recent SMP have better performance and have solved the
>> performance problem? Your suggestion will be highly appriciated.
>>
>Most likely it is the overhead of moving data between CPUs.  Could be
>cache effects also.
>
>--
>John
 
        
        Interesting.
        Is it true that SMP 3.0 kernel is still single-threaded ?
        Since stcp/rtcp is io intensive, it's like you still only
        have one thread of execution most of the time (like UP),
        and in the mean time, you lost the UP cache efficiency
        as well when two processors alternately execute the
        kernel ??
 
 
        - Tina

--=_ORCL_2577328_0_1191971024111509_0
Content-Type:message/rfc822

Date: 23 Oct 97 17:14:48
From:"John S. Dyson <toor@dyson.iquest.net>" <owner-freebsd-smp@FreeBSD.ORG>
To:shieyuan@eecs.harvard.edu,(ShieYuan,Wang)
Subject:Re: SMP is slower when two processors are enabled
Cc:freebsd-smp@FreeBSD.ORG
Return-Path:<owner-freebsd-smp@FreeBSD.ORG>
Received:from maildrop.nc.com (proxy@nc.com [207.88.167.98])          by deimos.nc.com (8.8.4/8.8.4) with SMTP   id RAA23966 for <TYANG@nc.com>; Thu, 23 Oct 1997 17:50:15 -0700 (PDT)
Received:(from proxy@localhost) by maildrop.nc.com (8.6.12/8.6.9) id RAA17164 for <TYANG@nc.com>; Thu, 23 Oct 1997 17:48:43 -0700
Received:from sarip.sol.net(169.207.30.120) by sebastian.nc.com via smap (V2.0) id xma017160; Thu, 23 Oct 97 17:48:17 -0700
Received:from hub.freebsd.org (hub.FreeBSD.ORG [204.216.27.18])          by sarip.sol.net (8.8.7/8.8.4) with ESMTP   id TAA06466; Thu, 23 Oct 1997 19:36:29 -0500 (CDT)
Received:(from root@localhost)          by hub.freebsd.org (8.8.7/8.8.7) id RAA08117          for smp-outgoing; Thu, 23 Oct 1997 17:15:01 -0700 (PDT)          (envelope-from owner-freebsd-smp)
Received:from dyson.iquest.net (dyson.iquest.net [198.70.144.127])          by hub.freebsd.org (8.8.7/8.8.7) with ESMTP id RAA08096          for <freebsd-smp@FreeBSD.ORG>; Thu, 23 Oct 1997 17:14:56 -0700 (PDT)          (envelope-from toor@dyson.ique
Received:(from root@localhost) by dyson.iquest.net (8.8.7/8.8.5) id TAA01451; Thu, 23 Oct 1997 19:14:48 -0500 (EST)
Message-Id:<199710240014.TAA01451@dyson.iquest.net>
In-Reply-To:<199710232354.TAA27809@steward.eecs.harvard.edu> from ShieYuan Wang at "Oct 23, 97 07:54:21 pm"
Sender:owner-freebsd-smp@FreeBSD.ORG
X-Loop:FreeBSD.org
Precedence: bulk
MIME-Version: 1.0
Content-Type:text/plain; charset=US-ASCII
Content-Transfer-Encoding:7bit

ShieYuan Wang said:
> protocol stack to transfer unlimited data on the same local host. If I
> only enable one processor, the throughput can be 35 MB/sec. However if
> I enable the other processor hoping to get better throughput, I only get
> 12 MB/sec.
> 
> I know that locking overhead between multiple processors may be a lot. But
> I never thought that it could be so high. Could anyone give me some clues
> about this bad performance? My SMP 3.0 was installed about 6 months ago.
> Does the more recent SMP have better performance and have solved the 
> performance problem? Your suggestion will be highly appriciated.
> 
Most likely it is the overhead of moving data between CPUs.  Could be
cache effects also.

-- 
John
dyson@freebsd.org
jdyson@nc.com

--=_ORCL_2577328_0_1191971024111509_0--




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199710241714.KAA12913>