See previous note to AOLserver listserv as I just committed the code...
Might be some stuff of interest that would make sense to move back into
the core. I'd also like to see modifications made to Tcl to make it much
easier to plug in alternate allocators (Hoard, tcmalloc, etc.) where
using those allocators makes the most sense. I think, as you mentioned,
the current Tcl threaded allocator is optimized to be a high performance
malloc, which is achieved by all but eliminating lock contention, but at
the cost of higher memory usage.
If the exploratory work Jim and did showed anything, it's that it isn't
very easy, or wise, to try and make the current allocator into a "one
size fits all" solution - probably better to instead use the particular
allocator which is best suited for your particular needs (high
performance, low memory usage, etc.)
- n
[EMAIL PROTECTED] wrote:
I'm of course interested in seeing whatever variants of the threaded malloc
that are done. The original was also provided (with thanks) by the folks at
AOL, designed on the original mods to threaded Tcl and high perf malloc for
it.
If this is all just refined improvement on the same, then it should go
straight back into the Tcl core.
Jeff Hobbs, The Tcl Guy, http://www.ActiveState.com/
Jim Davidson wrote:
Folks,
Nate and I spent a lot of time -- frankly, far too much time --
chasing this down a few months back. We ended up completely re-
writing the multi-threaded allocator with features to reclaim memory,
fix stats counters which in the current code are broken and lie, and
added an "alloc" command with all sorts of fun introspection
features, and some monitoring adp pages. Nate will post it for folks
to try if interested -- consider it interesting but experimental :)
-Jim
On Jun 2, 2006, at 2:37 PM, Guan Yang wrote:
Thanks,
What values would you suggest for NBUCKETS and MAXALLOC?
And looking at tclThreadAlloc.c, I'm not quite sure how to make the
spacing closer.
Guan
On 2 Jun 2006, at 19:28 , Jeff Rogers wrote:
I think one difference is in the tcl version itself; the latest
versions use a fast multithreaded allocator that allocates on
average slightly more than double the amount requestes, while
earlier versions may have been more frugal. You can cut your
memory usage by a third or more by increasing the number of
buckets defined in tclThreadAlloc.c and making the spacing closer;
there will be a bit more constant overhead and the allocator would
run microscopically slower. You may also want to look into the
ttrace module which makes all the tcl interpreters lazily load
their code instead of loading everything up front which can be a
significant savings but I have had very little luck with it (I
think the openacs code does too many fancy things to work well
with it).
-J
Guan Yang wrote:
Hi,
A couple of weeks ago I moved a number of servers to AOLserver
4.0.10 from various versions of 3.4 because of the 10^9 seconds
problem. After the move servers that were preivously using
100-120 MB of memory are now using around 300 MB to 400 MB each.
I've tried to turn off or reduce adp and fastpath caches,
reducing maxthreds to the 40's level. Stacksize is 500k. We have
used more or less the same config files, with a few changes like
the new nsopenssl configuration format.
Has anyone else had experiences with memory usage after switching
to AOLserver 4.0, and is there any easy way to reduce it?
Thanks,
Guan
--
AOLserver - http://www.aolserver.com/
To Remove yourself from this list, simply send an email to <[EMAIL PROTECTED]>
with the
body of "SIGNOFF AOLSERVER" in the email message. You can leave the Subject:
field of your email blank.
--
AOLserver - http://www.aolserver.com/
To Remove yourself from this list, simply send an email to <[EMAIL PROTECTED]>
with the
body of "SIGNOFF AOLSERVER" in the email message. You can leave the Subject:
field of your email blank.
