Mass storage bandwidth

Don Speck speck at cit-vax.arpa
Mon Dec 16 10:26:04 AEST 1985


>   ... not impressed with the speed of the drive [GCR CacheTape] -
>   seems only 10-20% faster than my 125 ips vacuum column 1600 bpi
>   Cipher 920 on the Unibus.

    Probably your disks are the bottleneck - even if they're Eagles.
Setting the 'rotational delay' parameter (tunefs -d) correctly helps
some; for an Emulex controller on a 750, 7ms seems about right (but
verify this on your own machine, since one millisecond less than the
optimum is the least-optimum value, half the speed of the optimum).
Write something like this (rotdelay.c):

	char buf[4096]; 	/* Your filesystem block size here */
	main(argc,argv) int argc; char *argv[]; {
		int gap = 0;
		if (argc > 1) sscanf(argv[1], "%d", &gap);
		if (gap*512 < -sizeof(buf)) lseek(0, 35*51*16384L, 0);
		while (read(0, buf, sizeof(buf)) == sizeof(buf))
			lseek(0, gap*512L, 1);
	}

Run it with the raw disk on stdin, with various arguments, until
you find one that maximizes transfers/second according to iostat.
For 4.2bsd, convert this to milliseconds:  1000*gap/(nsect*rps).
For Eagles, rps=66, nsect=48.  When correctly set, iostat will
report nearly 100 tps.	If you try this on an RA81, no value is
much bigger than any other; I think the track-switching time and
the "Unibus delay" (throttle) dominate over the rotation time.

<FLAME1>
I don't like tuning parameters.  They're usually very difficult
to set correctly, very easy to set incorrectly, poorly documented,
and so almost everyone takes the default - which in this case was
chosen to be just about the least-optimum value for Eagles.

<FLAME2>
Everybody loves to swap benchmark results, but nobody ever talks
about tuning.  Okay, let's call "rotdelay" a benchmark:
	4.2bsd	Vax/780     SI 9900	Eagle	    7.3 ms
	   "	   "	       "	CDC 9766    7.8 ms
	   "	Vax/750  Emulex SC7000	Eagle	    7.9 ms
	   "	   "	       "	Fuji 2298   8.1 ms
	Sun 2.0  Sun-2	 Xylogics 450	Eagle	   10.5 ms
	   "		     SCSI	71 MB	   12.5 ms
If I receive any "benchmark" results, I'll summarize after
Christmas.
		Don Speck	speck at cit-vax.arpa



More information about the Comp.unix.wizards mailing list