https://jira.whamcloud.com This file is an XML representation of an issue en-us 9.4.14 940014 05-12-2023 https://jira.whamcloud.com/browse/LU-35 Lustre <p>Despite the existence of the cfs_hash_t hash table class, the global hash table of portals_handle has its own fixed-bucket-count hash table implementation, using RCU for the lists of handles.</p> <p>This issue relates to a patch to be pushed to gerrit, which will:</p> <ul> <li>Remove the RCU support from the portals_handle hash table (RCU is less portable, requires extra complexity in code using portals_handle, and makes the portals_handle struct larger)</li> <li>Convert the portals_handle hash table to be a standard cfs_hash_t (reducing code size and getting advantage of hash resizing)</li> <li>Introduce per-cpu variables for the portals_handle cookie generation to reduce contention on a global lock when allocating cookies (potentially improving SMP performance)</li> </ul> LU-35

Make portals_handle use the generic cfs_hash_t

Improvement Minor Resolved Incomplete John Spray John Spray Wed, 5 Jan 2011 06:35:41 +0000 Wed, 17 Apr 2013 21:58:11 +0000 Wed, 17 Apr 2013 21:58:11 +0000 Lustre 2.0.0 Lustre 2.1.0 0 3 <p>See <a href="http://review.whamcloud.com/#change,167" class="external-link" target="_blank" rel="nofollow noopener">http://review.whamcloud.com/#change,167</a></p> <p>Changes to lustre_handle can have a significant impact on SMP performance, which was the original reason to add RCU to this structure (see <a href="http://bugzilla.lustre.org/show_bug.cgi?id=10706" class="external-link" target="_blank" rel="nofollow noopener">http://bugzilla.lustre.org/show_bug.cgi?id=10706</a> for the original problem that was resolved by this change). We need to validate this new patch by running it on a many-core SMP system under load. Liang should comment on the testing method and data analysis used during the development of the SMP scaling patches.</p> <p>Also, please remove the "#if !defined(HAVE_RCU)" compatibility code at the top of lustre/obdclass/lustre_handle.c as part of this patch, since it looks to be unnecessary with the removal of RCU usage.</p> <p>Andreas, it's cool to see you here, <img class="emoticon" src="https://jira.whamcloud.com/images/icons/emoticons/smile.png" height="16" width="16" align="absmiddle" alt="" border="0"/></p> <p>there are a few reasons we might want to change lustre_handle:</p> <ul> <li>some handles are not rarely changed, i.e: the most commonly used one is ldlm_lock handle, I'm not sure but probably RCU is not quite good choice for that case</li> <li>the RCU related stuff increased about 40 bytes for each handle, so we have more memory footprint.</li> <li>it would be good to have a few bits as "type" information of "cookie"(or have different handle types in different hash-table?), so we got less chance to screw up memory if remote peer passing a totally wrong handle to local which is a different data type. I'm not sure whether it's the thing we should consider about</li> </ul> <p>Anyway, I totally agree we do need to measure performance of the patch when we got environment...</p> <p>Cleaning up: this has been idle for &gt;2yrs.</p> Development Rank 1|hzvofj: Rank (Obsolete) 7777