From owner-freebsd-hackers Sat Oct 7 16:31:25 1995 Return-Path: owner-hackers Received: (from root@localhost) by freefall.freebsd.org (8.6.12/8.6.6) id QAA01357 for hackers-outgoing; Sat, 7 Oct 1995 16:31:25 -0700 Received: from Sysiphos (Sysiphos.MI.Uni-Koeln.DE [134.95.212.10]) by freefall.freebsd.org (8.6.12/8.6.6) with SMTP id QAA01350 for ; Sat, 7 Oct 1995 16:31:21 -0700 Received: by Sysiphos id AA12665 (5.67b/IDA-1.5 for hackers@freebsd.org); Sun, 8 Oct 1995 00:30:58 +0100 Message-Id: <199510072330.AA12665@Sysiphos> From: se@zpr.uni-koeln.de (Stefan Esser) Date: Sun, 8 Oct 1995 00:30:58 +0100 In-Reply-To: Bruce Evans "Re: VLB Disk Controllers" (Oct 8, 9:07) X-Mailer: Mail User's Shell (7.2.6 alpha(2) 7/9/95) To: Bruce Evans Subject: Re: VLB Disk Controllers Cc: hackers@freebsd.org Sender: owner-hackers@freebsd.org Precedence: bulk On Oct 8, 9:07, Bruce Evans wrote: } Subject: Re: VLB Disk Controllers } >Yup, but your whole machine is affected by a PIO device. } } That's mainly because IDE PIO is slow, not because it's PIO. The } overhead for SCSI DMA is not insignificant. For the BT455C, which } has a DMA speed of up to 40MB/sec (I doubt that my bus can keep up } with this), the overhead for reading from a Grand Prix at 5MB/sec } seems to be about 27%. This is hard to measure accurately. I } measured it by running a hog process in the background: on an idle } system it took 7.58 sec (real) and while transferring it took 10.42 } sec. 7.58/10.42 = 73%. Well, the machine is more severely affected by a PIO device, than the CPU load indicates. The CPU must be able to respond to data becoming available within a very short time, or data may be lost. (Hmm, maybe the larger buffers of current IDE drives make this a non- issue ???) Anyway, just a data point: http://www-itg.lbl.gov/ISS/performance.ps contains SCSI controller performance data, and they get 24.5MB/s simultaneously reading from 6 drives connected to 3 NCR 53c810 controllers in pairs. The system is reported to be 70% idle under that load (it's a Pentium 100 with Triton chip set). Another system seems to saturate its bus at 20MB/s with no CPU idle time left. But since there is 25% idle time reported reading 19MB/s from 5 drives, I guess it's just a problem arbitrating for the PCI bus in that system's SiS chip set. Well, if the above 24.5MB/s at 30% CPU load number is true, then you should be able to get some 60 to 70MB/s before other processes are locked out ... STefan -- Stefan Esser, Zentrum fuer Paralleles Rechnen Tel: +49 221 4706021 Universitaet zu Koeln, Weyertal 80, 50931 Koeln FAX: +49 221 4705160 ============================================================================== http://www.zpr.uni-koeln.de/staff/esser/esser.html