Re: XY and Memory Weirdness

[scarter@xxxxxxxx (Stephen A. Carter)]

In article <199701252002.PAA20179@xxxxxxxx>, Robert Holmgren wrote:
>Interesting. I know little about caches, but opinions sure seem to differ re:
>L2. Some say, that if you need to turn L2 off in order to install OS/2, then
>leave it off. Others say that problems arise only when L2 is configured for
>read-through rather than write-back caching. Still others swear by L2
>and claim huge performance benefits when they raise L2 cache size from,
>say, 256 to 512K. Whereas, there's a rumor that Intel's gonna dump L2 in the
>next generation processor. I see no difference at all on my machine, enabled or
>disabled, though I can't claim any scientific "tests". All very confusing. I
>conclude that L2 has very machine-specific effects. You have no hesitation about
>the broad prescription to disable it?

Well, I didn't until you asked me straight out. :-)

I've been told by my local guru to always turn it off under Warp, and
I'm sure I've read it in documents from IBM, but I've just done some
deeper digging and it looks like that advice is just plain wrong.

For sure, there are a lot of APARs and technical notes that describe
problems corrected by disabling L2 caching -- searching the closed
APAR database at http://service2.boulder.ibm.com/pspapar/ for "(L2 |
external) NEAR cach*" turned up a dozen and a half -- but apparently
it's not IBM's official recommendation to disable the external cache
in every case.

IBM's document WARPPERF at the Boulder ftp site says this:

> You want processors that can be upgraded and those
that support instruction caching sometimes referred to as Level 1 or Level 2 cache.
> There really is no optimum level instruction cache size. More instruction cache is
> usually better than less. The only real determinant should be costs.

And in his "Official OS/2 FAQ"
(http://www.secant.com/sipples/faqlist.htm/), IBM employee Timothy
Sipples says pretty much the same thing:

> If you need to save money, save it by getting a system with a less
> powerful processor....Also, the more processor cache memory you can
> afford in your PC, the better.

For what it's worth, I also found this on page 204 of _Inside OS/2
Warp Version 3_ from NRP (ISBN 1-56205-378-7):

> OS/2 Warp's disk performance can sometimes be improved by turning
> off the disk adapter's on-board cache. This is because double
> caching of the same data can cause a reduction in overall
> performance.

Also, IBM's PC Co. BBS (919/517-0001) has a couple of third-party
utilities for benchmarking cache performance, but I haven't tried them
out yet (486cache.zip, cachchk4.zip, and cachechk.zip).

My previous post was my first to the XyWrite list in maybe two years.
I'm very sorry to all -- especially Harmon Seaver -- that it was less
than accurate.

And just to keep this post marginally Xy-related: I've never been
able to figure out how to get XyWrite III+ to make direct use of
memory over 1 MB. My workaround has been to make a 2- or 3-MB RAM
disk (plenty for my usual needs) and make that XyWrite's default
directory (with "df DR="), but I'd love to learn about a more elegant
technique, if there is one.

--
Stephen A. Carter      High-Tech Information Center Ltd., Nagoya
                   Nagoya, Japan
http://www.hticn.com