Hi Aleksey (and anyone else interested!)
Do you have any further comments on these changes? With the recent
discussion over Tony's work on removing duplicate locale data I was
wondering if it would be ok to step up to this stable release if icu4j
and deal with performance issues "in the wild"? It seems to me that
while we stall upgrading to a release version of this library it's
performance will never be closely examined, and as such will continue to
put the upgrade on hold indefinitely.
I would suggest we move to icu4j 3.8.1 and begin to feed back
performance queries to the ICU team. Are there objections?
Regards,
Oliver
Oliver Deakin wrote:
Hi Aleksey,
Thanks for performance testing the changes. It's great that the Dacapo
benchmark has not been affected by this change, but I agree that the
small SPEC degradation is a concern. However, I would argue that we
should still upgrade to the new version. At this point I value bug
fixes and stability over performance, and IMHO moving to a proper
release version of icu4j is preferable to using the current mid
development cycle build we have in SVN, even if there is a slight
degradation in one of the benchmarks.
I would personally suggest that we still move to the new version of
icu4j after M4 and address the performance issues as an ongoing task
with the ICU developers.
Regards,
Oliver
Aleksey Shipilev wrote:
Hi, Oliver!
This change:
a. does not affect composite Dacapo performance
b. degrades SPECjbb2005 performance for 1-2%.
Taking (b) into the account, I would like to vote for deferring
immediate moving to ICU 3.8.1. We can revisit this one more time if
there's a way to beat the degradation.
Thanks,
Aleksey.
On Dec 14, 2007 3:53 PM, Oliver Deakin <[EMAIL PROTECTED]>
wrote:
Thanks for offering to help Aleksey, that's great!
Ive opened HARMONY-5313 and attached a script and a patch to be
applied.
Please run the script before applying the patch, as one file needs
to be
moved before it is patched. Once it's applied, if you run the
fetch-depends target you should see that the new ICU4J jars are
downloaded. Then if you just rebuild you should be ready to run with
the
new version.
Thanks for the help!
Oliver
Aleksey Shipilev wrote:
Thanks, Oliver!
Performance data on microtests is looking good, however I wonder what
impact it has on DRLVM and large benchmarks. Haven't you filed JIRA
for this issue? If I'll have exact steps to build configuration with
new ICU, I could check it's performance before committing the changes
to the trunk.
Aleksey.
On Dec 13, 2007 8:08 PM, Oliver Deakin
<[EMAIL PROTECTED]> wrote:
Thanks Aleksey.
I have run the JUnit tests with the IBM VME and they pass without any
new failures. I have also run the encoding/decoding test in
HARMONY-3709
a number of times - in general, ICU4J 3.8.1 performs in that test as
well or slightly better than the ICU4J 3.8 jar we are currently
using. I
have attached the results of the latest run I made [1].
Regards,
Oliver
[1]
DECODING:
Small Input:
Decoding: GB18030 , 1000000 times
Current ICU4J Milliseconds: 890.0
ICU4J 3.8.1 Milliseconds: 593.0
Small Input:
Decoding: ISO-8859-1 , 1000000 times
Current ICU4J Milliseconds: 328.0
ICU4J 3.8.1 Milliseconds: 328.0
Small Input:
Decoding: UTF-8 , 1000000 times
Current ICU4J Milliseconds: 344.0
ICU4J 3.8.1 Milliseconds: 360.0
Large Input:
Decoding: GB18030 , 1000 times
Current ICU4J Milliseconds: 2110.0
ICU4J 3.8.1 Milliseconds: 1968.0
Large Input:
Decoding: ISO-8859-1 , 1000 times
Current ICU4J Milliseconds: 157.0
ICU4J 3.8.1 Milliseconds: 156.0
Large Input:
Decoding: UTF-8 , 1000 times
Current ICU4J Milliseconds: 735.0
ICU4J 3.8.1 Milliseconds: 719.0
ENCODING:
Small Input:
Encoding: GB18030 , 1000000 times
Current ICU4J Milliseconds: 969.0
ICU4J 3.8.1 Milliseconds: 1063.0
Small Input:
Encoding: ISO-8859-1 , 1000000 times
Current ICU4J Milliseconds: 344.0
ICU4J 3.8.1 Milliseconds: 344.0
Small Input:
Encoding: UTF-8 , 1000000 times
Current ICU4J Milliseconds: 359.0
ICU4J 3.8.1 Milliseconds: 359.0
Large Input:
Encoding: GB18030 , 1000 times
Current ICU4J Milliseconds: 7407.0
ICU4J 3.8.1 Milliseconds: 7297.0
Large Input:
Encoding: ISO-8859-1 , 1000 times
Current ICU4J Milliseconds: 219.0
ICU4J 3.8.1 Milliseconds: 219.0
Large Input:
Encoding: UTF-8 , 1000 times
Current ICU4J Milliseconds: 625.0
ICU4J 3.8.1 Milliseconds: 610.0
Aleksey Shipilev wrote:
Good news, Oliver!
Anyway, basing on our previous experiences with ICU changes, we
might
first try how Harmony performs with new ICU onboard before making
this
change default. That's not only JUnit and other validation
suites, but
performance too (say, Dacapo and other benchmarks performance).
Thanks,
Aleksey.
On Dec 13, 2007 2:53 PM, Oliver Deakin
<[EMAIL PROTECTED]> wrote:
ICU 3.8.1 has just been released, so Id like to propose that
after M4 we
upgrade to this release, add it to the fetch-depends target and
remove
the "homemade" ICU4J jar we have stored in SVN.
Im running the tests with the new version now to make sure there
are no
regressions and will be happy to make the required changes when
the time
comes.
Objections?
Regards,
Oliver
--
Oliver Deakin
Unless stated otherwise above:
IBM United Kingdom Limited - Registered in England and Wales
with number 741598.
Registered office: PO Box 41, North Harbour, Portsmouth,
Hampshire PO6 3AU
--
Oliver Deakin
Unless stated otherwise above:
IBM United Kingdom Limited - Registered in England and Wales with
number 741598.
Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire
PO6 3AU
--
Oliver Deakin
Unless stated otherwise above:
IBM United Kingdom Limited - Registered in England and Wales with
number 741598.
Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire
PO6 3AU
--
Oliver Deakin
Unless stated otherwise above:
IBM United Kingdom Limited - Registered in England and Wales with number 741598.
Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire PO6 3AU
