From richmc@entropy.ucsc.edu Wed Oct 29 00:25:18 1997 Received: from entropy.ucsc.edu for richmc@entropy.ucsc.edu by www.ccl.net (8.8.3/950822.1) id UAA08674; Tue, 28 Oct 1997 20:26:53 -0500 (EST) Received: (qmail 16452 invoked by uid 500); 29 Oct 1997 01:26:42 -0000 From: "Rich McClellan" Message-Id: <9710281726.ZM16450@entropy.ucsc.edu> Date: Tue, 28 Oct 1997 17:26:42 -0800 In-Reply-To: Bob Zellmer "Re: CCL:PCI vs. SCSI and Linux" (Oct 28, 18:24) References: <199710282324.SAA29117@phosphorus.mps.ohio-state.edu> X-Mailer: Z-Mail (3.2.1 24feb96 Caldera) To: Bob Zellmer Subject: Re: CCL:PCI vs. SCSI and Linux Cc: chemistry@www.ccl.net Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii On Oct 28, 18:24, Bob Zellmer wrote: > Subject: Re: CCL:PCI vs. SCSI and Linux > > Rich: > > Actually, your % of CPU went down, but this doesn't necessarily mean the > actual CPU time went down. Your CPU time could have increased by running the > utility while the spent doing I/O decreased (faster I/O) and you would still > see a decrease in the % CPU time spent in the I/O. I will assume however Hi, I agree that faster I/O will reduce the time spent by the CPU negotiating I/O, but the magnitude of difference between the two results is too large, me thinks, for this explanation to account for the drastic drop in CPU usage. To test this I just ran the benchmarks again with the time command, with and without the EIDE drive tuned, and got the following results. (The fields real, user and sys are the results from the time command.) EIDE drive tuned [16:43:16] entropy:/usr/local/bonnie# time ./bonnie -d /opt -m entropy -------Sequential Output-------- ---Sequential Input-- --Random-- -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks--- Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU entropy 100 3309 53.9 3071 15.1 1282 12.2 2672 30.4 3189 11.9 35.0 1.9 real 5m46.626s user 0m20.140s sys 0m29.170s EIDE drive NOT tuned [16:51:30] entropy:/usr/local/bonnie# time ./bonnie -d /opt -m entropy -------Sequential Output-------- ---Sequential Input-- --Random-- -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks--- Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU entropy 100 2503 77.1 2867 32.9 1192 39.1 2206 75.3 2991 76.4 34.8 5.3 real 6m15.828s user 0m39.370s sys 1m45.230s With the drive tuned, the CPU spent roughly 20 + 29 seconds to handle the benchmark. Without tuning, the CPU spent 39 + 95 seconds to do the same task. The relative change in I/O does not match the change in CPU usage, so I conclude that the CPU time decreased with the drive tuned and, moreover, with bus mastering off the CPU must handle more of the I/O . > that actual CPU time also decreased. Perhaps these numbers are in your > benchmark file, which I haven't had the opportunity to look at just yet. > I've included my original post in case you lost it. If I'm missing something here, please enlighten me :-) Rich > Bob Zellmer > > Robert J. Zellmer > Department of Chemistry > The Ohio State University > Columbus, OH > zellmer@sodium.mps.ohio-state.edu >-- End of excerpt from Bob Zellmer Subject: CCL:PCI vs. SCSI and Linux I apologize for replying to this email so long after it was originally posted, but I have conducted some benchmarks on my system that you people may find interesting. My system: o PPRO 180 MHz (256 kb cache) o Buslogic BT-948 SCSI controller (fast, scsi-ii) o Conner fast, scsi ii disk (model no.CFP2107S, 256kb cache and spindle rotation rate of 7200 rpm) o Triton chipset for EIDE controller o Seagate EIDE model no. ST3780A (256 kb cache, 4500 rpm) o Running the Linux OS of course :-) Two factors that influenced my decision to install a SCSI controller and disk were the perceived higher I/O of the SCSI disks and the smaller fraction of CPU used during reads/writes. Results: The SCSI disk has throughput consistently higher than the EIDE disk (probably due to the higher rpm) BUT when the EIDE disk is tune with the hdparm utility to enable bus mastering and mulitple sector I/O, the throughout of the EIDE drive increases but more notable the CPU usage GOES DOWN DRASTICALLY. Have a look at the benchmark results. My conclusions: Go with SCSI (preferably fast, wide), but EIDE properly tuned is a cheap alternative. ***************************BENCHMARK RESULTS********************************* (Note that sda is the scsi disk and hda is the eide disk) On sda: -------Sequential Output-------- ---Sequential Input-- --Random-- -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks--- Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU entropy 100 4021 65.0 3848 21.9 1631 17.8 4758 57.0 4483 19.4 53.8 3.2 on hda: Note: drive NOT tuned with hdparm utility -------Sequential Output-------- ---Sequential Input-- --Random-- -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks--- Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU entropy 100 2642 82.4 2783 27.5 1140 37.7 2071 69.7 2589 59.7 32.3 4.7 on hda: Note: Drive tuned with hdparm utilities (-d1 -X34 -m32) -------Sequential Output-------- ---Sequential Input-- --Random-- -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks--- Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU entropy 100 3115 50.1 2750 12.4 1228 12.6 2557 29.0 2871 10.1 32.7 1.6 #################################### 150 Mb cache ############# SCSI drive -------Sequential Output-------- ---Sequential Input-- --Random-- -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks--- Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU entropy 150 4237 70.6 4241 22.7 1543 16.2 3167 37.3 4058 16.4 53.2 3.2 EIDE drive, hdparm on -------Sequential Output-------- ---Sequential Input-- --Random-- -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks--- Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU entropy 150 2794 44.2 2569 12.7 1179 11.9 2468 27.2 2758 10.1 28.0 1.3 ******************************************************************************* The benchmarks are called "bonnie". You can get them from any Linux mirror like Sunsite. I hope these are helpful, Rich -- Molecular Engineering Laboratory Department of Chemistry and Biochemistry University of California, Santa Cruz > -------This is added Automatically by the Software-------- > -- Original Sender Envelope Address: richmc@entropy.ucsc.edu > -- Original Sender From: Address: richmc@entropy.ucsc.edu > CHEMISTRY@www.ccl.net: Everybody | CHEMISTRY-REQUEST@www.ccl.net: Coordinator > MAILSERV@www.ccl.net: HELP CHEMISTRY or HELP SEARCH | Gopher: www.ccl.net 73 > Anon. ftp: www.ccl.net | CHEMISTRY-SEARCH@www.ccl.net -- archive search > Web: http://www.ccl.net/chemistry.html >-- End of excerpt from Rich McClellan -- Molecular Engineering Laboratory Department of Chemistry and Biochemistry University of California, Santa Cruz