jeremy@uunet.uu.net (Jeremy Webber) (05/03/89)
ames!mailrus!metasoft!alan@uunet.uu.net (Alan Epstein) writes:
4- how good is the performance in doing something like this?
Our hardware engineers, who use a CAD package on Sun 3/60's did some
performance comparisons of having a Sun shoebox on a 4MB Sun 3/60 vs a
diskless 4MB sun 3/60 vs diskless 8MB and 12MB.
The shoebox contained the root and swap partitions, but not the binaries for
the CAD package.
The result was:
The diskfull 4MB Sun actually ran slightly SLOWER than the diskless 4MB one
The diskless 8MB Sun ran twice as fast as the diskless 4MB one
The diskless 12MB Sun ran at the same speed the diskless 8MB Sun.
The timing was the real time taken to perform a fixed sequence of tasks
using a portion of the CAD package noted for its memory consumption. The
network had ample spare capacity. The server machine was a Sun 4/280,
which also had spare capacity.
Conclusions:
Sun shoeboxes aren't as fast as an Ethernet.
If you are short of memory, more memory is MUCH better than local disks. It
is also much cheaper.
Use the money you save on shoebox disks to buy more disk capacity on your
server.
As a result of the tests, we traded our shoeboxes for more memory. The
engineers are happier with performance on their Suns now (before we got
frequent complaints of poor respone).
If you don't have enough memory, you are wasting staff time.
Provisos:
The testing was not an exhaustive benchmark.
You must have a server with adequate capacity.
If you have enough memory, net traffic due to swapping should not be a worry.
Of course, net traffic due to other NFS load may still be, but that won't
be solved by putting your swap partition on a local disk.
Hope this helps,
-jeremy
--
--
Jeremy Webber ACSnet: jeremy@chook.ua.oz
Computer Science Department Internet: jeremy@chook.ua.oz.au
University of Adelaide Voicenet: +61 8 228 5763
GPO Box 498, Adelaide 5001 Papernet: +61 8 223 1206 (FAX)
Australia
scs@vax3.iti.org (Steve Simmons) (05/09/89)
munnari!chook.ua.oz.au!jeremy@uunet.uu.net (Jeremy Webber) writes: >Our hardware engineers, who use a CAD package on Sun 3/60's did some >performance comparisons of having a Sun shoebox on a 4MB Sun 3/60 vs a >diskless 4MB sun 3/60 vs diskless 8MB and 12MB. > >The shoebox contained the root and swap partitions, but not the binaries for >the CAD package. > >The result was: > The diskfull 4MB Sun actually ran slightly SLOWER than the diskless 4MB one > The diskless 8MB Sun ran twice as fast as the diskless 4MB one > The diskless 12MB Sun ran at the same speed the diskless 8MB Sun. > >Conclusions: > Sun shoeboxes aren't as fast as an Ethernet. > If you are short of memory, more memory is MUCH better than local disks. It > is also much cheaper. Jeremy did put some provisos and a good description of methods in his posting, but I'm afraid too many folks will look at results/conclusions and not apply them appropriately for the circumstances. Jeremy had lots of unused capacity on server. We did the same sort of tests, but under radically different conditions. Our servers supported 9-10 diskless clients apiece, all of which were used heavily for software development and testing of CAD/CAM software. If the server was heavily loaded and the client was doing a lot of paging/swapping, performance was better with local disk. Reason: one server with SMD disk can serve that disk to one client very quickly. One server with SMD disk can serve that disk to ten *simultaneous* clients very slowly. Also note that in many circumstances adding memory is not an economically viable option: shoeboxes can be moved to other machines. 3/50 add-on memory cannot. In the longer term, one must balance three choices: expand the memory on your 3/50s and commit to them for a long time; expand the local disk on your 3/50s with the view that disk will go elsewhere in the future; replace your 3/50s with 3/80s. Steve Simmons Just another midwestern boy scs@vax3.iti.org -- or -- ...!sharkey!itivax!scs "Hey...you *can* get here from here!"
weber@harvard.UUCP (Jeff Weber) (05/10/89)
munnari!chook.ua.oz.au!jeremy@uunet.uu.net (Jeremy Webber) writes: >Conclusions: > Sun shoeboxes aren't as fast as an Ethernet. > If you are short of memory, more memory is MUCH better than local disks. It > is also much cheaper. I agree with your conclusions but I think there's more to it...... A. We found that 3/50's were helped more than 3/60's for our large text&graphics pagination jobs. No explanation was found after minimal search. B. A Client Local Disk (CLD) won't impact performance much if the CLD's disk is slow. You want it be as close to memory speed as possible. Thus ours is 16ms access time and uses embedded SCSI to reduce propagation delays. C. A CLD won't impact performance much if the fileserver and net are pretty lightly loaded. It shows mucho bettero performance when the server is on its knees. D. A CLD did allow us to move some of the server's swap space to /usr space and we could also put more clients on a net with acceptable performance. E. At <$1300 a 91MB CLD WAS cheaper than a single 4MB memory upgrade. With our swap space on the fileserver at 25MB, a 75MB swap space on a fast CLD was nice. (Our paginated documents will run in the 100's of MB). F. For us, 8MB was enough for nice "interactive" performance. But "dedicated wait" or "batch performance" was helped by getting away from the usual Sun disk I/O bottle neck. Now we can have multiple users running multiple paginations without locking each other out. G. CLD's are most useful in a large network with heavily loaded clients. Clients that only do memo's and read the news obviously don't need/use a CLD. But that's just one man's opinion. Jeff Weber