ndd@duke.UUCP (03/27/87)
In article <1588@psivax.UUCP> woof@psivax.UUCP (Hal Schloss) writes: > >... Each Sun has a very small disk attached >to it with the minimum needed to boot UNIX and get NFS up on the air. They >each have a small /tmp filesystem and a reasonably sized swap area. These >are intended to speed up processing on the Suns... >-- > Hal Schloss > Pacesetter Systems Inc. > {sdcrdcf|ttidca|scgvaxd|nrcvax|jplpro|hoptoad|csun|quad1| > bellcore|logico|rdlvax|ihnp4}! psivax!woof This brings up an interesting question: is it faster to swap locally on one of Sun's 70 Meg disks over a SCSI interface, or over ethernet? I realize that this will depend on how loaded the ethernet and the server(s) are; does anyone have any ballpark figures? Ned Danieley ndd@duke
jeff@fluke.UUCP (04/11/87)
> ndd@duke.UUCP asks: > This brings up an interesting question: is it faster to swap locally on > one of Sun's 70 Meg disks over a SCSI interface, or over ethernet? > I realize that this will depend on how loaded the ethernet and the > server(s) are; does anyone have any ballpark figures? Some engineers at our site did these very tests a few months ago. Here is a summary of their results. I further recall hearing that a local disk used for a root filesystem makes a workstation FEEL much more responsive than using ND (Sun's current virtual disk block protocol). I assure you that a completely diskless Sun-2 can be a very lethargic workstation. Jeff Stearns John Fluke Mfg. Co, Inc. (206) 356-5064 {uw-beaver, decvax!microsoft, ucbvax!lbl-csam, allegra, sun}!fluke!jeff -------------------------------------------------------------------------- A few months ago, we ran some tests to get some rough measures on disk throughput. Each trial consisted of reading a quasi-random number of contiguous blocks (in the range 0-100) from a quasi-random place in the disk partition 1000 times. The same sequence of random selections was used for each trial. Tests ran during the night with otherwise unloaded clients, servers, and Ethernet. During daytime hours, a busy server and Ethernet will cause very noticable performance losses; we did not attempt to measure performance under these conditions. /dev/rsd0b: (this is a Sun-2 SCSI system, using an Adaptec ACB-4000 controller and a Micropolis 1304 drive) partition size: 23460 512-byte blocks # of trials: 143 # of reads per trial: 1000 bytes transferred per trial: 26.196 MB minimum trial elapsed time: 102.10 sec => 257 KB/sec overall maximum trial elapsed time: 110.43 sec => 237 KB/sec overall /dev/rsd1b: (this is the Sun-2 with a borrowed SCSI system, using an Adaptec ACB-4000 controller and a Miniscribe 8425 drive) partition size: 31280 512-byte blocks # of trials: 143 # of reads per trial: 1000 bytes transferred per trial: 26.314 MB minimum trial elapsed time: 142.56 sec => 185 KB/sec overall maximum trial elapsed time: 151.00 sec => 174 KB/sec overall /dev/rnd1: (this is a Sun-2 client's ND swap partition on a Sun-2 server) partition size: 33200 512-byte blocks # of trials: 142 # of reads per trial: 1000 bytes transferred per trial: 26.355 MB minimum trial elapsed time: 194.14 sec => 136 KB/sec overall maximum trial elapsed time: 232.43 sec => 113 KB/sec overall ======================== The following measurements were made in the same way on a Sun-3/75 during off-hours. This is an ND swap partition on a Sun-3 server. /dev/rnd1: (this is the Sun-3 client's ND swap partition on a Sun-3 server) partition size: 34040 512-byte blocks # of trials: 15 # of reads per trial: 1000 bytes transferred per trial: 26.318 MB minimum trial elapsed time: 114.43 sec => 230 KB/sec overall maximum trial elapsed time: 144.12 sec => 183 KB/sec overall We ran only 15 trials because the results were pretty consistent and the server was very lightly loaded. Most of the trials were 115-125 seconds. It's interesting to note that 230 KB/sec is 1.84 Megabits/sec, a substantial fraction of the ethernet 10 MB/sec theoretical maximum. Jeff Stearns 206-356-5064 new: jeff@tc.fluke.COM John Fluke Mfg. Co. old: uw-beaver!fluke!jeff P.O. Box C9090 Everett WA 98206 or: allegra!fluke!jeff