writing out disk cache

From: LA Walsh (law@sgi.com)
Date: Mon Nov 13 2000 - 16:52:39 EST

Another question that's been bugging me -- this is behavior that seems
identical in 2216/2217 and not related to my ealier performance degredation
post.

I run VMware. It runs w/144Mg and writes out a 153M suspend file when I
suspend it to disk. My system has a total of 512M, so the entire
suspend file gets written to the disk buffers pronto (often under 1 second).

But a 'sync' done afterwards can take anywhere from 20-40 seconds.
vmstat shows a maximum b/o rate of 700, with 200-500 being typical.

So, I know that the maximum write rate through the disk cache is
close to 10,000 blocks/second. So why when the disk cache of a
large file is 'sync'ed out, do I get such low b/o rates?

Two sample 'vmstat 5' outputs during a sync were:
 1 0 0 6292 13500 254572 165712 0 0 1 0 119 282 1 1
98
 2 0 0 6292 13444 254572 165716 0 0 0 702 279 534 0 2
98
 1 1 0 6292 13444 254572 165716 0 0 0 501 352 669 0 1
99
 0 1 0 6292 13444 254572 165716 0 0 0 520 372 697 0 2
97
 1 0 0 6292 13444 254572 165716 0 0 0 510 367 694 0 2
98
 0 1 0 6292 13444 254572 165716 0 0 0 694 379 715 0 2
98
 1 0 1 6292 13444 254572 165716 0 0 0 618 391 964 0 2
98
 0 1 1 6292 13444 254572 165716 0 0 0 441 302 765 0 1
98
 0 0 0 6292 13496 254572 165716 0 0 0 63 180 355 1 1
99
 0 0 0 6292 13496 254572 165716 0 0 0 0 103 195 0 1
99
----and
 0 0 0 6228 18836 246036 167824 0 0 0 0 232 563 6 13
82
 0 1 0 6228 18784 246036 167824 0 0 0 506 175 489 2 1
97
 1 0 0 6228 18780 246036 167824 0 0 0 292 305 647 0 1
99
 0 1 0 6228 18780 246036 167824 0 0 0 253 285 602 0 1
99
 0 1 0 6228 18780 246036 167824 0 0 0 226 289 612 0 1
99
 1 0 0 6228 18832 246036 167824 0 0 0 157 199 406 0 1
99
 0 0 0 6228 18832 246036 167824 0 0 0 0 101 240 1 1
99 

---
	Another oddity -- If you add up the rates in the 2nd example, and multiply
the average rate by 5, you get around 5200 blocks written out (for a 152M
file).
Note that a du on it shows it to use 155352, so it isn't that it is sparse.

	Is vmstat an unreliable measure?  The above tests were on a 2216 system.

-l

--
L A Walsh                        | Trust Technology, Core Linux, SGI
law@sgi.com                      | Voice/Vmail: (650) 933-5338

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

This archive was generated by hypermail 2b29 : Wed Nov 15 2000 - 21:00:24 EST