Обсуждение: Data write speed
Folks, I have a new system with an Adaptec 2200S RAID controller. I've been testing some massive data transformations on it, and data write speed seems to top out at 3mb/second .... which seems slow to me for this kind of hardware. As it is, I have RAM and CPU sitting idle while they wait for the disks to finish. Thoughts, anyone? -Josh
Josh Berkus kirjutas T, 18.02.2003 kell 23:45: > Folks, > > I have a new system with an Adaptec 2200S RAID controller. RAID what (0,1,1+0,5,...) ? > I've been > testing some massive data transformations on it, and data write speed > seems to top out at 3mb/second .... which seems slow to me for this > kind of hardware. As it is, I have RAM and CPU sitting idle while they > wait for the disks to finish. > > Thoughts, anyone? How have you tested it ? What OS ? --------------- Hannu
Josh Berkus wrote: > Folks, > > I have a new system with an Adaptec 2200S RAID controller. I've been > testing some massive data transformations on it, and data write speed > seems to top out at 3mb/second .... which seems slow to me for this > kind of hardware. As it is, I have RAM and CPU sitting idle while they > wait for the disks to finish. > > Thoughts, anyone? That does seem low. What rate do you get with software RAID (of the same type, of course) to the same disks (might have to be through a standard SCSI controller to be meaningful) with roughly the same disks/channel distribution? My experience with hardware RAID (at least with the hardware available a few years ago) versus software RAID is that software RAID is almost always going to be faster because RAID speed seems to be very dependent on the speed of the RAID controller's CPU. And computer systems usually have a processor that's significantly faster than the processor on a hardware RAID controller. It's rare that an application will be as CPU intensive as it is I/O intensive (in particular, there are relatively few applications that will be burning CPU at the same time they're waiting for I/O to complete), so the faster you can get your I/O completed, the higher the overall throughput will be even if you have to use some CPU to do the I/O. That may have changed some since CPUs now are much faster than they used to be, even on hardware RAID controllers, but to me that just means that you can build a larger RAID system before saturating the CPU. The Adaptec 2200S has a 100MHz CPU. That's pretty weak. The typical modern desktop system has a factor of 20 more CPU power than that. A software RAID setup would have no trouble blowing the 2200S out of the water, especially if the OS is able to make use of features such as tagged queueing. Since the 2200S has a JBOD mode, you might consider testing a software RAID setup across that, just to see how much of a difference doing the RAID calculations on the host system makes. -- Kevin Brown kevin@sysexperts.com
Kevin, > That does seem low. What rate do you get with software RAID (of the > same type, of course) to the same disks (might have to be through a > standard SCSI controller to be meaningful) with roughly the same > disks/channel distribution? Unfortunately, I can't do this without blowing away all of my current software setup. I noticed the performance problem *after* installing Linux, PostgreSQL, Apache, PHP, pam_auth, NIS, and my database .... and, of course, customized builds of the above. > The Adaptec 2200S has a 100MHz CPU. That's pretty weak. The typical > modern desktop system has a factor of 20 more CPU power than that. A > software RAID setup would have no trouble blowing the 2200S out of > the > water, especially if the OS is able to make use of features such as > tagged queueing. > > Since the 2200S has a JBOD mode, you might consider testing a > software > RAID setup across that, just to see how much of a difference doing > the > RAID calculations on the host system makes. This isn't something I can do without blowing away my current software setup, hey? -Josh
Josh Berkus wrote: <snip> > This isn't something I can do without blowing away my current software > setup, hey? Hi Josh, Any chance that it's a hardware problem causing a SCSI level slowdown? i.e. termination accidentally turned on for a device that's not at the end of a SCSI chain, or perhaps using a cable that's not of the right spec? Stuff that will still let it work, but at a reduced rate. Maybe even a setting in the Adaptec controller's BIOS? Might be worth looking at dmesg and seeing what it reports the SCSI interface speed to be. :-) Regards and best wishes, Justin Clift > -Josh -- "My grandfather once told me that there are two kinds of people: those who work and those who take the credit. He told me to try to be in the first group; there was less competition there." - Indira Gandhi
Murthy, > You could get mondo (http://www.microwerks.net/~hugo/), then backup your > system to CDs and restore it with the new filesystem layout. You might want > to do these backups as a matter of course? Thanks for the suggestion. The problem isn't backup media ... we have a DLT drive ... the problem is time. This particular application is already about 4 weeks behind schedule because of various hardware problems. At some point, Kevin Brown and I will take a weekend to swap the postgres files to a spare disk, and re-format the data array as pass-through Linux RAID. And this is the last time I leave it up to the company sysadmin to buy hardware for a database server, even with explicit instructions ... "Yes, I saw which one you wanted, but the 2200S was on sale!" -- Josh Berkus Aglio Database Solutions San Francisco
On Thu, 27 Feb 2003, Josh Berkus wrote: > Murthy, > > > You could get mondo (http://www.microwerks.net/~hugo/), then backup your > > system to CDs and restore it with the new filesystem layout. You might want > > to do these backups as a matter of course? > > Thanks for the suggestion. The problem isn't backup media ... we have a DLT > drive ... the problem is time. This particular application is already about > 4 weeks behind schedule because of various hardware problems. At some point, > Kevin Brown and I will take a weekend to swap the postgres files to a spare > disk, and re-format the data array as pass-through Linux RAID. > > And this is the last time I leave it up to the company sysadmin to buy > hardware for a database server, even with explicit instructions ... "Yes, I > saw which one you wanted, but the 2200S was on sale!" I still remember going round and round with a hardware engineer who was extolling the adaptec AIC 133 controller as a great raid controller. I finally made him test it instead of just reading the pamphlet that came with it... Needless to say, it couldn't hold it's own against a straight symbios UW card running linux software RAID. He's the same guy who speced my workstation with no AGP slot in it. The week before he was laid off. Talk about bad timing... :-(