[comp.arch] Write-back controllers

davidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr) (05/01/91)

  After all the discussion of write back and write thru controllers, in
flavors dumb, smart, and brilliant, I got a chance to do an actual test
on one the other day, and I thought I'd mention the results. Note that
these are the results of a single test, under unfavorable conditions. I
expect to get more formal testing done in the future.

  The test system was a Dell 433E (EISA) system with Ultrastor 22
controller, with only 512k cache on board. The usual disk tests showed
very little difference between the WD1007 emulation mode and the EISA
mode (with device driver installed) and write back enabled. However...

  The system is being used for software development of a LARGE
proprietary system. The system is configured such that changing certain
things makes it desirable to recompile everything because of multiple
dependencies. Using the WD1007 or Ultrastor 22 in normal mode, this took
26 minutes clock time. When the new driver was installed and the write
back enabled the time dropped to 15 minutes. CPU time is 13.2 minutes.

  Whatever the drawbacks of write back, for this application the factor
of two improvement in performance makes them irrelevant. Installation of
a UPS, combined with regular backups, put the reliability at an
acceptable level.

  To say I was surprised was an understatement. I have seen good gains
by going to a caching controller, but nothing like 2:1.
-- 
bill davidsen	(davidsen@crdos1.crd.GE.COM -or- uunet!crdgw1!crdos1!davidsen)
        "Most of the VAX instructions are in microcode,
         but halt and no-op are in hardware for efficiency"

jerry@talos.npri.com (Jerry Gitomer) (05/01/91)

davidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr) writes:


:  After all the discussion of write back and write thru controllers, in
:flavors dumb, smart, and brilliant, I got a chance to do an actual test
:on one the other day, and I thought I'd mention the results. Note that
:these are the results of a single test, under unfavorable conditions. I
:expect to get more formal testing done in the future.

	[configuration details omitted]

:  The system is being used for software development of a LARGE
:proprietary system. The system is configured such that changing certain
:things makes it desirable to recompile everything because of multiple
:dependencies. Using the WD1007 or Ultrastor 22 in normal mode, this took
:26 minutes clock time. When the new driver was installed and the write
:back enabled the time dropped to 15 minutes. CPU time is 13.2 minutes.

:  Whatever the drawbacks of write back, for this application the factor
:of two improvement in performance makes them irrelevant. Installation of
:a UPS, combined with regular backups, put the reliability at an
:acceptable level.

	In a development environment I too would opt for performance, but 
	I hope that my bank uses write-thru.

-- 
Jerry Gitomer at National Political Resources Inc, Alexandria, VA USA
I am apolitical, have no resources, and speak only for myself.
Ma Bell (703)683-9090  (UUCP:  ...uunet!uupsi!npri6!jerry )