dan@msdc.UUCP (11/27/83)
I had my interest tweaked by all the questions on disk interleaving, so I thought I'd apply some brute force to complement the subtle logic applied by others. The results were somewhat surprising, so I thought y'all might be interested [note the appropriate use of the second person plural pronoun :-)]. I also suspect that not many folks have completely idle systems to fiddle with, either. We are running 4.1bsd on 750s with RL02s, RA80s and RA81s in various mixes. I ran various combinations of interleave factors for the various disks reading and writing a 2MB (2^20B) file. I have no faith in the "writing" numbers, since I was copying a file from another drive with different interleave, so I've included summaries of only the reading elapsed times (the user and system times were nearly identical in all cases). The systems were quiescent except for cron waking up to do nothing and a uucp logging in every 5 minutes or so. These might skew a single trial, so I did the reads twice. First, here's the script I used to make it happen: ------ for x in 3 5 7 9 11 13 15 17; do for y in 15 30 60 120 217 434; do echo "m n: $x $y" fsck -s$y:$x /dev/ra0f >/dev/null mount /dev/ra0f /b2 echo -n " Writing: " time sh -c "dd if=/usr/tmp/bigfile of=/b2/test bs=10240 >/dev/null 2>&1" echo -n " Reading: " time sh -c "dd if=/b2/test of=/dev/null bs=10240 >/dev/null 2>&1" echo -n " Reading: " time sh -c "dd if=/b2/test of=/dev/null bs=10240 >/dev/null 2>&1" rm -f /b2/test umount /dev/ra0f done done ------ And now, here are the numbers. One column is for each device type, with the interleave factors and average elapsed read time for a 2MB file (using 'dd' with bs=10240). They are presented in order of increasing average time. ------RL02------ ------RA80------ ------RA81------ m n Times m n Times m n Times 3 10 23 24 7 30 18 18 11 208 23 22 4 10 25 26 7 217 18 18 11 728 23 22 3 20 26 26 7 434 18 18 11 364 23 23 3 40 29 28 7 15 18 19 10 728 24 23 2 10 29 29 7 120 19 18 11 52 24 24 2 5 31 29 5 30 22 21 11 104 24 24 3 5 31 31 7 60 21 22 12 728 24 25 5 10 31 31 5 60 22 22 11 26 25 25 2 20 31 32 5 15 20 25 13 728 25 25 2 40 33 32 5 120 22 23 12 26 26 26 4 20 33 33 5 217 22 23 13 208 26 26 4 40 34 34 9 434 23 22 13 364 26 26 6 10 36 35 5 434 23 23 13 104 27 26 5 20 41 40 9 15 23 23 10 26 27 28 4 5 42 41 9 30 24 23 13 26 28 27 5 40 42 41 9 120 24 23 15 728 28 28 6 40 48 48 9 60 24 24 13 52 29 28 6 20 49 48 9 217 27 23 15 104 29 29 3 60 26 25 15 208 30 28 3 120 26 25 15 364 29 29 3 434 25 26 15 26 29 30 3 30 26 26 9 52 30 30 3 217 26 26 9 26 31 31 3 15 27 26 9 104 31 31 11 120 27 27 9 208 31 31 11 30 27 28 9 364 31 31 11 217 27 28 9 728 31 32 11 15 27 29 15 52 31 32 11 60 28 28 17 26 32 31 11 434 32 27 17 728 32 31 17 60 31 30 17 364 32 32 13 15 32 31 17 208 33 32 13 60 32 31 17 104 34 33 13 434 32 31 7 52 39 37 15 120 32 32 7 364 39 38 13 30 33 32 7 26 39 39 13 120 32 34 7 208 39 39 13 217 33 33 7 728 39 39 15 217 33 33 5 26 40 39 15 434 33 33 5 104 40 39 15 60 33 34 5 208 39 40 17 15 34 33 5 728 39 40 17 120 35 35 7 104 40 39 15 30 36 36 17 52 36 43 17 30 37 36 3 26 40 40 17 217 37 36 5 52 40 40 15 15 37 38 5 364 40 40 17 434 38 37 3 104 41 40 3 208 41 40 3 364 41 41 3 728 41 41 3 52 42 41 I welcome comments on the interpretation of these numbers. I'd especially like to know why the RA81s are so much slower than the RA80s. Dan Forsyth (allegra!akgua!msdc!dan, decvax!duke!mcnc!msdc!dan) Medical Systems Development Corp, Atlanta