From owner-freebsd-scsi  Fri Sep 20 21:30:28 1996
Return-Path: owner-freebsd-scsi
Received: (from root@localhost)
          by freefall.freebsd.org (8.7.5/8.7.3) id VAA09101
          for freebsd-scsi-outgoing; Fri, 20 Sep 1996 21:30:28 -0700 (PDT)
Received: from silvia.HIP.Berkeley.EDU (silvia.HIP.Berkeley.EDU [136.152.64.181])
          by freefall.freebsd.org (8.7.5/8.7.3) with ESMTP id VAA09059;
          Fri, 20 Sep 1996 21:30:21 -0700 (PDT)
Received: (from asami@localhost) by silvia.HIP.Berkeley.EDU (8.7.6/8.6.9) id VAA14425; Fri, 20 Sep 1996 21:29:42 -0700 (PDT)
Date: Fri, 20 Sep 1996 21:29:42 -0700 (PDT)
Message-Id: <199609210429.VAA14425@silvia.HIP.Berkeley.EDU>
To: michaelv@MindBender.serv.net
CC: freebsd-scsi@freebsd.org, freebsd-current@freebsd.org,
        freebsd-isp@freebsd.org
In-reply-to: <199609210411.VAA20862@MindBender.serv.net> (michaelv@MindBender.serv.net)
Subject: Re: Streamlogic RAID array benchmarks
From: asami@freebsd.org (Satoshi Asami)
Sender: owner-freebsd-scsi@freebsd.org
X-Loop: FreeBSD.org
Precedence: bulk

 * >Yeah, it's a P5 (133MHz).  We got pretty much the same result with the
 * >P6 (200MHz) too (which is kinda surprising, given that their memory
 * >system is so much slower).
 * 
 * How is that surprising?  The SCSI controller lives on the other side
 * of the bus, and does the bus-mastering irrespective of the CPU.  The
 * CPU does not do bcopies for bus-mastering SCSI transfers.

It does, because this is a bufferred I/O.  We were measuring this
through the filesystem, remember?

An interesting sidenote of this is that when we were using the same
P5-133 with the regular (rep/movsl) copyin/copyout, we got only about
20MB/s.  We changed it to the Pentium-optimized copy, and managed to
push it up close to 30MB/s.

The slow bcopy can move about 40MB/s, the fast one 80MB/s.  However,
the P6-200, despite its 45MB/s bcopy speed, gives us the same 30MB/s
through the filesystem.

(For those you like math, 45MB/s is 22.2ns/byte, and 30MB/s is 33.3ns/B,
 and given that the max transfer rate on of 33MHz PCI is 132MB/s, or
 7.5 ns/B, this is really close to the limit (22.2 + 7.5 = 29.7 =~
 33.3)....)

Satoshi

P.S. Doing parallel reads from raw devices, I could get about 65MB/s
     on the P5...haven't tried that on the P6.