From owner-freebsd-questions Thu Dec 23 12:51:49 1999 Delivered-To: freebsd-questions@freebsd.org Received: from corinth.bossig.com (corinth.bossig.com [208.26.239.66]) by hub.freebsd.org (Postfix) with ESMTP id 335891575B for ; Thu, 23 Dec 1999 12:51:15 -0800 (PST) (envelope-from kstewart@3-cities.com) Received: from 3-cities.com (kenn1223.bossig.com [208.26.241.223]) by corinth.bossig.com (Rockliffe SMTPRA 3.4.5) with ESMTP id ; Thu, 23 Dec 1999 12:58:12 -0800 Message-ID: <38628B94.AAA938BF@3-cities.com> Date: Thu, 23 Dec 1999 12:52:36 -0800 From: Kent Stewart Organization: Columbia Basin Virtual Community Project X-Mailer: Mozilla 4.7 [en] (WinNT; U) X-Accept-Language: en MIME-Version: 1.0 To: "Alexey N. Dokuchaev" Cc: freebsd-questions@FreeBSD.ORG Subject: Re: when is it safe to use the 0xa0ffa0ff disk flags? References: Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Sender: owner-freebsd-questions@FreeBSD.ORG Precedence: bulk X-Loop: FreeBSD.ORG "Alexey N. Dokuchaev" wrote: > > Hi! > > > CDC (now out of the business) had a hard drive called a Hydra. It was > > a full height 8" HD that ran 9.6MB/s in 1988. It was also very > > expensive at $250,000 US. Cray striped them together to create what > > they called a DD40. It had 16 Hydras and had a total of 20GB. By > > striping 4 together, Cray would get 20MB/s when they were hung on the > > 100MB/s data channel. When our benchnmark was run on a system with the > > Hydras for main storage, the throughput basically doubled. The Cray > > XM/P had 2x the throughput of the CDC-990. The 990 was a dinosaur and > > represented the end of big iron. It filled the room where as the Cray > > looked like the what was left of the volcano in "Close Encounters of > > the Third Kind". I think the 8" FH HD's were all replaced by HH 5.25" > > HD's, which looked pretty puny mounted in the center of the Hydra's > > bay. The cooling requirements probably dropped 10KVA :). The effect of > > the Hydra on the benchmark was enormous. This the benchmark that I > > told you about writing behind was so important. The benchmark ran > > almost twice as fast when write behind caching was used. I think > > SOFTUPDATES produce an imporovement that helps for similar reasons. > > Is there anywhere where I can read more about it? If you are asking about the benchmark, the answer is no. Setting up for a benchmark is similar to setting up a race car for a time trial. You want it to run as fast as possible with a special setting. Those settings may not be appropriate when it is released to the users. A race car setup for a time trial is probably dangerous when you change the balance by loading a full load of fuel onboard. The programs were all designed around licensing commercial nuclear power reactors in the US and the actual runs were considered proprietary. Some of the programs depended on 60-64 bit numbers. One of the programs had to be able to and/or a 60-bit integer with FORTRAN. The floating point number extremes ran from 10^(+/-)2048 or 4096. The old computers ran from ~10^(+/-)308 The mix of jobs was an average day as defined by accounting data for a year. They depended on the batch job capability of the computers. The benchmark started out by building the programs and then running them. The benchmark shell script added new jobs to the batch queue every so often. Some of them modified a module and then build a new program. The benchmark also included 100 simulated users doing various sorts of things on line. The user simulation ran on the Sun front ends that were on all Cray's at that point in time. The users were a constant minor irritation to the system. The Cray was eventually replaced by 3 really small DEC Alphas. Before that happened each group had a number of HP Xterms and a couple of HP 9000 servers to run their jobs interactively. The production jobs were run on the Cray and then on the Alphas. By small I mean the desk where the Cray monitors were hung was large enough to include all three Alpha and the 5 or 6 - 20GB drives attached to each. Each Alpha had several times the memory of the Cray. Fast memory was really expensive in 1988. I think the Cray needed something like 230KVA of cooling running at all times. I think, at this point in time, my PC's have more memory than they did. Each Alpha (<200MHz) had 1.5 times the throughput of the Cray. Setting up the Alpha's cost less than the maintenance on the Cray was running. The benchmark was considered obsolete because a single large computer in the future was considered highly unlikely and it was archived on to magnetic tape(s), which were probably recycled when I was retired. If you are running in a similar environment, the things we learned still apply. You never have a fast enough file system. Write behind caching to your HD's can be 50% of your system throughput. Turn it off and you may run half as many jobs in the same amount of time. If you are swapping too much, you don't have enough memory. If you aren't swapping a little bit, you spent too much money on memory. From then on it is a matter of fine tuning your system. BTW, the user percived speed of the system will be gauged on the amount of time required for the prompt to reappear after they press the "enter" key :). I haven't seen a real comparison of SOFTUPDATE in a mix of environments. If you are trying to figure out how important it is, try running a number of your jobs with timing turned on, turn SOFTUPDATE on and run them again. You can redirect input on most programs and develop scripts based on that ability. You could time it by using the shell script to append the redirected the output of "time" to a file and automate the data collection. The kind of things most people are using FreeBSD for is different than the stuff we were doing and the tuning of the machines would be different. A buildworld could represent a piece of what was going on. A kernel build is probably closer. When I do a buildworld, I leave Seti@Home running in the background. The cpu is always running at 100% until I do an install. There are no other users that could sneak things in on me. On a batch oriented computer, for example, you didn't do things like run a large database. Web and mail servers would also be running on a different computer. Since our benchmark was a purchase requirement, it was run against a stop watch. The batch system had to be idle in x hours. When I run setiathome, which has no real I/O requirements, on various systems, I find it is mostly memory limited and PC-66 memory slows it down tremendously. When I replaced a Celeron 433a with an Celeron 300a@450, the average time per workunit dropped from a little over 54,000 secs to a little under 34,000 seconds. The average on the 300a@450 is coming up on 100 WU's. This works out pretty close to the ratio of PC-66/PC-100 access times. I saw a ~15% improvement with the 433 by just replacing the PC-66 memory with PC-100 memory. Setiathome is running much faster; however, the buildworld times only dropped ~300 seconds, i.e., ~2160u to ~1890u. Memory speed is important but not as much as it was to setiathome. Depending on how many gaussians they are seeing a P-III 450 or the Celeron 300a@450 jumps into the lead on average processing speed. The P-III 450 is running Windows 2000 Server and the Celeron is running FreeBSD. I have a 2% variation between motherboards and much less than that between OS'es. Setiathome running on a P-II 400 on a SuperMicro P6SBA mb is running 2% faster than it does on a P-II 400 on an ABIT BX6 rev 2 mb. All of the BIOS setting are on defaults. Kent > > ./danfe -- Kent Stewart Richland, WA mailto:kstewart@3-cities.com http://www.3-cities.com/~kstewart/index.html FreeBSD News http://daily.daemonnews.org/ SETI(Search for Extraterrestrial Intelligence) @ HOME http://setiathome.ssl.berkeley.edu/ Hunting Archibald Stewart, b 1802 in Ballymena, Antrim Co., NIR http://www.3-cities.com/~kstewart/genealogy/archibald_stewart.html To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-questions" in the body of the message