pshimizu@fa2.so-net.ne.jp
Sat, 28 Nov 1998 00:42:22 +0900
Hi, Mr. Martin,
In article <199811270749.IAA14161@keksy.mchp.siemens.de>
martin.kahlert@mchp.siemens.de writes:
>> Hi,
>> i will own an alpha box in about 2 weaks or so and i want to do
>> heavy numerical calculation (analog circuit simulation) on it.
>>
It is more difficult for alpha to treat the problem like circuit
simulation, because it uses extensive sparce matrices and
alpha (21164A) has only 64 TLB entry and 5 MAF.
But it is very interesting problem to treat the sparce matrices and I
had a plan to make the task as my undergraduate student thesis.
Unfortunatly, there are few student in these years and I could not
do the plan yet.
>> I would like to help improving the performance of Alpha software, too.
>> It seems to me, that Mr. Goto is T H E guru of alpha performance.
>> Hope he's listening.
>> Is there any introduction paper/www-site, where i can learn about
>> his methods? (e.g. cache optimization/scheduling...) perhaps with
>> some C-examples (or does this only make sense in assembler?)
If you can read Japanese, I recommend you to read our mailing list
Linux-Alpha-JP. In the list, I and Mr. Goto and some other people
discussed about that theme for a long time.
The MAF and the TLB usage and load scheduling will be the key point
for the good performance (of cource, the blocking is basic). These
scheduling strongly depend on the high level structure of the program.
And you can control from FORTRAN or C.
The other minor scheduling will be covered with Out-Of-Order feature
of 21264.
Then there is a hope that we can use C or FORTRAN to retrieve full
performance from Alpha when released the 21264 for public(it means
low cost).
But I myself is not comfortable with 21264, because it does not
improve the peak floating point performance from 21164. The 21364
will use the same core with 21264 and then we cannot expect
outperformed performance with these chips.
I hope Samsung will enhance the 21164A with good price/performance.
Regards,
Naohiko Shimizu
School of Engr., Tokai University
http://shimizu-lab.et.u-tokai.ac.jp/indexe.html
This archive was generated by hypermail 2.0b3 on Sat Nov 28 1998 - 08:53:21 EST