stevens@hsi.UUCP (12/02/83)
Regarding all the notes recently concerning the m&n values for
mkfs on the new disks (RA80 and RA81), the original work on this
topic was done in 1979 by Walt Lazear and Charles Muir at the Air Force
Data Services Center and presented at the January 80 Usenix conference
in Boulder. Their work was done under Version 6 but the approach is
still valid and consists of:
- Measure the CPU speed in blocks/sec (512-byte blocks are used
throughout). To do this execute the program
main()
{
char buff[512]; /* 1024 under BSD */
while (read(0, buff, 512 or 1024) > 0)
;
}
using some large file as input (more than 2Mb). Calculate the
system's blocks/sec (BPS) by dividing the file size in 512-byte blocks
by the SYSTEM time to read. (Make sure the file does not have
"holes" in it - that all the bytes are really there.)
Note that the interleave factor used to measure this makes no difference
as you are using the system time, not the clock time.
- Calculate the number of blocks/sec that the disk generates.
This is the number of 512-byte blocks per track. I think you'd
better use all the blocks on the track, even if some are not
used by the software (alternate sector handling by the controller);
for example, use 52 for the RA81, not 51.
- Calculate the next integer greater than
disk BPS / CPU BPS
and you have the interleave factor, m.
As Ed Bryant at Simon Fraser found out by tests, the m value is
much more important than the n value.
Some values that I have measured and some values that appeared in the
original paper for system speeds are:
System BPS Unix version
------ --- ------------
IBM PC 113 Venix/86, XT disk, I measured
11/34 119 Version 6, from paper
11/70 210 Version 6, from paper
11/70 190 Version 7, I measured
11/750 385 4.1 BSD, without FPA, I measured
11/750 410 4.1 BSD, with FPA, I measured
For the RA81 on a 750 without the FPA, m = 3120 / 52 = 9, which is
just what Ed Bryant measured and I got the same value in some
identical tests. (3120 = 60 revs/sec * 52 blocks/track.)
Similarly, for an RL02 on a 750 without the FPA, m = 800 / 385 = 3,
which is what I measured on an actual RL02.
The difference in Unix versions on the 11/70 numbers above are due
to the additional overhead in V7 from V6. The difference in the two
11/750 values must be due to the speed increase in the integer multiply
and divide when the FPA is present, since I don't think the kernel uses
floating point anywhere in the block i/o system or in any of the disk
drivers.
In summary, the 4.1 BSD mkfs manual page that says to use m=3 all the time
is wrong. Calculate your own values ! (But you don't need to run
all the timing tests - use the formula above.)
I'd be very interested if anyone with a 730 or 780 running 4.1 BSD could
calculate and post the system speeds (BPS) with and without the FPA.
Richard Stevens
Health Systems International, New Haven, CT
{ decvax | hao | seismo | sdcsvax } ! kpno ! hsi ! stevens
ihnp4 ! hsi ! stevens
p.s. - I believe all of this is irrelevant under 4.2 BSD.wls@astrovax.UUCP (William L. Sebok) (12/03/83)
> - Calculate the next integer greater than > disk BPS / CPU BPS > and you have the interleave factor, m. > As Ed Bryant at Simon Fraser found out by tests, the m value is > much more important than the n value. > In summary, the 4.1 BSD mkfs manual page that says to use m=3 all the time > is wrong. Calculate your own values ! (But you don't need to run > all the timing tests - use the formula above.) Still one needs to run timing tests to establish the CPU BPS figure, especially if your system is not one of the types on the above list. But as said above once one knows this number for one disk one should be able to compute it for the other disks on the system. I believe that 4.2 may do something like this. One the other hand, one thing still worries me, that different disk drivers may have different disk overheads, especially if they are very different, like, say, one is a massbus disk and the other a unibus disk. Anybody have any comments on that? -- Bill Sebok Princeton University, Astrophysics {allegra,akgua,burl,cbosgd,decvax,ihnp4,kpno,princeton}!astrovax!wls
bernerus@chalmers.UUCP (Christer Bernerus) (12/07/83)
Thank's for your info about m & n value calculation. I ran your program to check our system's speed. We're running a VAX 11/780 without FPA and 4.1BSD. The figure i got is 589 BPS which i think is reasonable. It would be interesting to know if anybody else has made your test on a machine similar to ours, with & without FPA. If one determines the m factor based on this configuration and a RM05 disk you'll find that m should be about 3.25 which implies that the interleave factor should be 4, not 3. Perhaps 3 is adequate if a FPA is used.