Bug #4818
closedgitbuilder: apparently building Ceph without tcmalloc
0%
Description
Users have reported a couple times that the "next" branch is being built without tcmalloc enabled. mikedawson installed the quantal packages on raring and when he tries to do a heap stats on the monitor, he gets the "tcmalloc not enabled, can't use heap profiler commands" output. But when I look at http://gitbuilder.sepia.ceph.com/gitbuilder-quantal-deb-amd64/log.cgi?log=ea54e6603f9c322d38f9f664dee92997b8a4bd4f it very clearly indicates that's using tcmalloc.
Figuring this out is important as not using tcmalloc makes our memory usage absolutely intolerable. Is this a gitbuilder issue that our release packages won't hit? Is something broken in the makefile? How did it happen?
Updated by Anonymous about 11 years ago
- Status changed from New to In Progress
For both debian and rpms, --with-tcmalloc is the default. It looks like it's setting the flags appropriately, and libtcmalloc is found. The Binaries that are produced are just not linked with libtcmalloc. Still looking.
Updated by Dan Mick about 11 years ago
Found it:
debian/rules explicitly disallows it because debian/control doesn't contain Build-Depends: and libgoogle-perftools-dev on the same line anymore.
This happened with
8f5c665744e58d6d51a1e86de55c1399f51cc1c3 debian: Run wrap-and-sort from devscripts Signed-off-by: Andras Elso <elso.andras@gmail.com>
I have a fix which seems to work for a local workspace that uses perl to implement a multiline grep:
-ifeq ($(shell grep -c -E '^Build-Depends:(.*,)?[[:space:]]*libgoogle-perftools-dev([[:space:]]|,|$$)' debian/control),0) +ifeq ($(shell perl -n -e 'BEGIN{$$/=""}' -e '/Build-Depends:.*libgoogle-perftools-dev/ && exit 1' debian/control),0)
Updated by Dan Mick about 11 years ago
- Assignee changed from Anonymous to Dan Mick
Updated by Anonymous about 11 years ago
Hi Dan -
I missed that you were working on this. I think the solution is to just delete that check since lenny is no longer a supported distribution, and all the others have tcmalloc.
There still appears to be a problem with the rpms, although it probably something completely different.
Updated by Dan Mick about 11 years ago
That's entirely valid too; I just wanted the builds to be available for testing ASAP, but I wondered if we even needed that. I'll let you take this back and complete it for the RPMs if you wish.
Updated by Sage Weil about 11 years ago
- Status changed from In Progress to Resolved
Updated by Dan Mick about 11 years ago
- Assignee changed from Dan Mick to Anonymous