keith@imagen.UUCP (Keith Rich) (02/16/90)
We recently upgraded our file servers and we are getting slower performance than we had expected. I'd like advice from anyone who has already been through this, or who can suggest what I might try to tune. We replaced six DSP80A/DSP90s with 500 MB disks, with one DN4000 with four 697 MB disks. Most of our workstations are DN3010s with 4 MB of memory and the DSP80As had 3 MB of memory. These systems run SR9.7 for two reasons. We have yet to convert some important programs to run correctly on SR10.2 and we know we will need more than 4 MB of memory to run SR10.2 efficiently. The DN4000 has 12 MB of memory and is running SR10.2 because SR9.7 will only suport two disks. The file system performs well directly on the DN4000, but poorly when serving files for the SR9.7 DN3010s. The degradation is perhaps a factor of 2.5, so I can't just forget it. I suspect that the bottleneck is some sort of interface between SR9.7 and SR10.2, but I'm not sure what or why. The performance acts this way even when I'm only using a single disk, so it isn't simply too many disks on one server. If the ring net were the bottleneck, then I not sure why the DSPs seem better than the DN4000. We are considering the following annoying choice. Upgrade another DN3010 to a DN4000, put two of the 697 MB disks on it, and downgrade both DN4000s to SR9.7. I'm reluctant to do this because it seems to mean that I cannot migrate to SR10.2. Instead, I will have to cut everything over at once (if ever). So, I fear that taking this step will simply lock me into SR9.7 forever. What other (better?) choices have I missed? Where should I look to find out more about how to tune things up? I also had to deal with another annoying problem when I moved my files from my SR9.7 disks to my SR10.2 disks. I discovered that using "cpt from to -pdt -sacl" preserved the Aegis ACLs, but lost the Unix permissions. Using "find ... | cpio -p ..." preserved the Unix permissions, but lost the Aegis ACLs. So I used cpt and then wrote a program to recover the Unix permissions. It worked like "cd oldtree; find . -print | modtree oldtree newtree". Has anyone found a simpler, more direct way to do this? Thanks (in advance), Keith Rich (keith@imagen.com)
lampi@stb.uucp (Michael Lampi) (02/20/90)
The Apollo filesystem under SR9.7 (and probably still under SR10.2) experiences a marked reduction in performance as one mounts additional logical disk volumes on a particular node. The reason for this behavior was explained to me as having to do with the fact that the relationship of UID's to volumes is nearly nonexistant, and the OS has to do a lot of searching to find a file amongst each volume. However, once the object (file) has been opened, access to that file is very fast. For example, if you have a floppy disk mounted along with a local winchester, the OS will take a look at the floppy when you are trying to locate a file. Of course, if the filesystem information of the floppy happens to be in RAM, then this access will be much faster than if it has to be re-read from the diskette. This is why you will hear/see the floppy being accessed from time to time when you haven't done anything to it for a long time. (Yes, I know that paging also occurs if you've just written something to the floppy, but this is after you've just mounted it and done nothing else.) I suggest that you consider striping and cylinder stacking your DN-4000 disks together, making them 1 or two logical volumes. With your dual controllers and 10.2 OS, you should get pretty good performance. Michael Lampi (213) 782-7888 Fax (213) 782-7927 MDL Corporation PO Box 745, Torrance, CA 90508