ok, all issues have been resolved and the fixes are in SystemDS' main.
The MSVM OOM was due a missing Xmn parameter, the naive bayes number
came from 878s down to 46s, the MSVM number from 655s down to 109s, and
we now added the missing glmPredict builtin function as well.
I'll now start a full fresh run of the perftest suite.
Regards,
Matthias
On 10/10/2021 4:47 PM, Matthias Boehm wrote:
just a quick update - yesterday, we ran the perftest from 8MB to 80GB
for all algorithms (and configurations) in the groups binomial,
multinomial, and regression. Overall it looks pretty good, except the
following issues, which we will address in the next days:
* MSVM 8GB ran OOM on both binomial and multinomial
* builtin glmPredict is still missing
* Performance issues for MSVM 80GB multinomial (85s w/o intercept but
655s w/ intercept) and NaiveBayes 80GB nomial (878s although non-iterative)
Regards,
Matthias
On 10/8/2021 12:55 PM, Matthias Boehm wrote:
ok, tonight I'll kickoff the perftest suite and see how it goes. In
the next week or two we should then restrict ourselves to bug fixes,
and hold off major features that are in the critical path (i.e.,
enabled by default).
Regards,
Matthias
On 9/13/2021 10:44 PM, Baunsgaard, Sebastian wrote:
Hi all,
+1 for the plan.
With regards to the performance suite,
I think it would be nice with a standard short 1 hour version as well.
best regards
Sebastian
________________________________
From: Matthias Boehm <[email protected]>
Sent: Monday, September 13, 2021 10:26:25 PM
To: [email protected]
Subject: [DISCUSS] SystemDS 2.2 Release / Upgrades
Hi all,
the original plan was to release SystemDS 2.1 as the last release for
JDK 8 and Spark 2.x. However, our main branch accumulated quite a number
of critical fixes and performance improvements. Accordingly, I'd like to
propose releasing SystemDS 2.2 again for JDK 8 and Spark 2.4 in the next
weeks and then immediately switch to JDK 11 and then release SystemDS
2.3. Any comments or opinions?
I would volunteer as release manager (never done it before) so we spread
the knowledge how to run the release process a bit. Of course, I would
appreciate help from previous release managers and people who
contributed to the automation of the release.
Finally, we recently revived the performance test suite so we can run a
battery of algorithm-level tests at different scales before the release.
I would make a case for running this for local/GPU, Spark, and federated
with scales that allow to complete it in ~1-2 weeks.
Regards,
Matthias
