Messages in this thread | | | Date | Wed, 25 Mar 1998 21:10:19 +0100 (CET) | From | Andreas Kostyrka <> | Subject | Re: Cheetah vs. UDMA: Bonnie says UDMA is faster! why? |
| |
On Wed, 25 Mar 1998, Mark Lehrer wrote:
> > Hello! > > I almost hate to put this on the linux-kernel list, but the newsgroups > were of literally no help... > > I have an NCR/Symbios 875 UW scsi controller and Cheetah 4GB hard > disk. I also have a 3GB UDMA drive (WD 33200). The UDMA is / and > the Cheetah is /home. This system is primarily a web server. > > I am disappointed in what Bonnie has to say about performance... it > claims that the UDMA edges out the Cheetah. Benchmarks are not really all. Basically the UDMA is probably ``faster'' for a workstation setting, but once you have multitasking/multiusing, than SCSI pays off. Basically I've had here a server with an old slow SCSI disc, and to do the ppl here something good, I decided to replace it with a fast new EIDE drive. (SCSI is awful expensive, AND the hype is that EIDE drives with DMA are also ok.) The funny thing is, that the new drive was in fact 2-3 times as fast at IO/rate (depending upon the exact benchmark), even the random access marks were higher for the EIDE disc. The CPU usage wasn't that much higher, so I put the production system on the EIDE disc. The nightmare came on the next Monday, where the system screached to a halt when two users started to work. (The old slow hdd worked quite OK for five users, whereby the program, fvwm, SQL backend etc. all run on the box.) I've tested it again, and again, but on paper the EIDE hdd was fast. Still it couldn't tolerate the multiuser load. After a week or so I moved the stuff to a two SCSI hdd system, where the system disc is actually a newer hdd (still about 30% slower than the EIDE drive), and the data is on the older one, that is about 2-3 times slower than the EIDE drive. After this little move, everybody is happy, and as I've plenty of RAM in the box, the users don't even notice that I'm compiling a kernel in the background :)
Morale of the history: SCSI may measure slower than EIDE, but test it in a real server environment, and you'll know why you are paying the money. > > Below are the boot messages: it claims at the end that it is running > "SLOW WIDE" instead of fast-20... that might be the entire problem. The > controller comes up and says fast-20 and this is the only device on the > chain so I can't blame it on an old device sharing the bus. If you are using the BSD derived driver, you have to setup the driver after booting.
Basically SCSI always transmits commands async with 5MB/s, and only the data transfers are fast. This way, you can have a slow old device on the bus, and especially if it can detach, you shouldn't have any performance impact. > > My question: is anyone else running this configuration? What can I do > to get performance up to where it should be? Does the "syncronous > transfers" section in the kernel config cause this or is it more likely > to be a jumper setting on the drive? Read the following file: /usr/src/linux/drivers/scsi/README.ncr53c8xx > > I have the drive on the very end of the SCSI cable; could that make a > difference? Nope. The bus should probably be terminated, etc. but that's it.
Andreas
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu
| |