<html><head><meta http-equiv="Content-Type" content="text/html charset=iso-8859-1"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space; "><div><div>On Apr 26, 2013, at 9:39 AM, Rajat Garg <<a href="mailto:garg_rajat@hotmail.com">garg_rajat@hotmail.com</a>> wrote:</div><blockquote type="cite"><div class="hmmessage" style="font-size: 12pt; font-family: Calibri; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: 2; text-align: -webkit-auto; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-size-adjust: auto; -webkit-text-stroke-width: 0px; "><div dir="ltr"><div dir="ltr"> I am using jemalloc 3.1.0, compiled with gcc 4.1.2 on CentOS release 5.6 for a compute intensive multithreaded application where bulk of allocations falls in 8-byte, 16-byte (stats below). We are seeing very high runtime in tcache_alloc_small_hard(), arena_tcache_fill_small() and arena_bin_malloc_hard(). The 8 and 16 bytes allocations happen in very large number. We probably want the allocations to come form tcache_alloc_easy() to not hit the locks and take less runtime. The allocations are mostly temporary in nature -- especially 8-byte ones -- so some number of 8-byte allocations are done, then they are freed and then alloc/free process repeated.<span class="Apple-converted-space"> </span><br><br>The runs use 8 threads (so 48 arenas and all other jemalloc settings are default) and are on an Intel Xeon server with 12-cores (no hyperthreading and no other user so no contention with other user; the peak memory in application is ~6.5GB and server has 128GB memory so no memory shortage/swapping etc.); the jemalloc output at the end of run is given below. As we can see, pretty much 8 and 16 byte bins are overloaded.</div></div></div></blockquote><div><br></div>Only 8 of the 48 arenas are being used in your test case. That said, the extra 40 are lazily initialized, so there's no need to change settings.</div><div><br><blockquote type="cite"><div class="hmmessage" style="font-size: 12pt; font-family: Calibri; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: 2; text-align: -webkit-auto; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-size-adjust: auto; -webkit-text-stroke-width: 0px; "><div dir="ltr"><div dir="ltr">A run with TCACHE_NSLOTS_SMALL_MAX=10000, LG_TCACHE_MAXCLASS_DEFAULT=10, changing small bins to hold only upto 224 bytes (NBINS=15), and hardcoding<span class="Apple-converted-space"> </span><br>tcache_bin_info[i].ncached_max = TCACHE_NSLOTS_SMALL_MAX in tcache.c file improves runtime by about 16% (for overall application runtime) but increases peak memory from 6.5GB to 6.7GB. We want the runtime to at least improve by another 10% without preferably any increase in peak memory (so even 6.5GB to 6.7GB is not desirable).<span class="Apple-converted-space"> </span><br>Any suggestions on what changes to jemalloc settings to try?<br></div></div></div></blockquote><br></div><div>I don't think the TCACHE_NSLOTS_SMALL_MAX setting is having as much effect for your test case as you might expect. Even with the setting of 10000, tcache is limited to 1002 and 504 regions for 8- and 16-byte regions, respectively. If you want to further increase the tcache region count, you will need to either increase the logical page size, or modify bin_info_run_size_calc() to use a larger initial min_run_size.</div><br><div>Jason</div></body></html>