Date: Sun, 7 Sep 1997 17:11:10 +0930 From: Greg Lehey <grog@lemis.com> To: Simon Shapiro <Shimon@i-connect.net> Cc: FreeBSD Chat <chat@FreeBSD.ORG> Subject: Re: lousy disk perf. under cpu load (was IDE vs SCSI) Message-ID: <19970907171110.27847@lemis.com> In-Reply-To: <XFMail.970906235822.Shimon@i-Connect.Net>; from Simon Shapiro on Sat, Sep 06, 1997 at 11:58:22PM -0700 References: <199709070512.AAA00465@dyson.iquest.net> <XFMail.970906235822.Shimon@i-Connect.Net>
next in thread | previous in thread | raw e-mail | index | archive | help
(moved to -chat) On Sat, Sep 06, 1997 at 11:58:22PM -0700, Simon Shapiro wrote: > > Hi "John S. Dyson"; On 07-Sep-97 you wrote: > > What you are all running into is the fact that, for a smoothly run system, > you need a certain balance between I/O, memory (capacity & bandwidth, > both), and CPU. > > If you compare a 14" disk pack from the mid-seventies (SMDE) to a disk > drive of today, you see that capacity climbed nicely (about 30x), > performance has barely moved (I am NOT talking about 5MB Shugart 5.25"!), > being that a good SMDE drive could do about 5MB/sec or even better, while > CPU's jumped almost 300x! Well, I can't remember the performance of the mid-70s, but my recollection of the performance in the early 80s on, say, a 3330 clone was that these drives had 30 sectors (2 spares) per track, and they ran at 3600 rpm. Since they weren't buffered, that gives a maximum data transfer to the channel of about 860 kB/s. Average positioning was round the 30 to 35 ms mark. At the time, I was working for Tandem. We noticed a puzzling behaviour: our new flagship model TXP, whose CPU was about 100% faster than the previous NonStop II, read data off these disks at almost exactly double the speed of the NonStop II. Why? We finally figured out that we were reading 4 kB blocks (the maximum our disk controllers would allow), doing some processing, and then issuing the next read. Unfortunately, on the NonStop II this took such a long time that the head had passed the next block before it issued the read. In other words, the "high performance" TXP managed to read a 4 kB block every revolution, and the NonStop II needed two revolutions per 4 kB block. This may sound funny until you look at the data transfer rates involved. On the TXP, it was 240 kB per second, on the NonStop II it was 120. Raw disk rate. > BTW, the architecture (in realizable terms) of the typical disk controller > has not changed at all since the IBM-PC, and even the best caching > SCSI controller is primitive when compared to a 1963 IBM mainframe, and not > much faster. Again, I can't agree in the slightest. I *do* have the technical doc for a 2311 drive floating around somewhere in the shed, and they were unbelievably primitive. > The young ones amoung you, I have a challenge for you (old folks like me > can think no more :-): You're not the only old fart around here. > Don't bellyache about disk performance. It will just get worst > (at 50%/year, last I checked). > > Come up with a new I/O ARCHITECTURE. Totally new. The first one to come > up with a truely random storage device with capacity of 1GB and performance > of (sustained) 100MB/Sec Sounds like a 1 GB RAM to me. Still cheaper per byte than any disk made up to about 5 years ago. > will make more money than BG thought exists. Not if BG has anything to say in it. Greg
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?19970907171110.27847>