As said on http://jemalloc.net/:
jemalloc is a general purpose malloc(3) implementation that emphasizes fragmentation avoidance and scalable concurrency support.
Some Sphinx and Manticore Search users prefer jemalloc over malloc and say it allows to save some RAM. I’ve tested how jemalloc affects resource consumption and response time for 2 weeks on 3 Manticore clusters (se03/03-2, se04/04-2, se05/05-2) each of which:
- was getting the same amount of traffic
- was hosting about the same amount of data
- had 2 Manticore servers fully identical in terms of hardware, software and with 100% identical indexes and even load balancing between the servers
On se03 and se03-2 I enabled jemalloc and disabled THP (transparent huge pages) and restarted Manticore on se03-2
On se04-2 I just restarted Manticore, I didn’t touch se04. se04/04-2 is a control group.
On se05-2 I enabled jemalloc, but didn’t disable THP
Here’s what I’ve got in 2 weeks:
RSS RAM consumption:
Latency dynamics (I used moving average(response time)):
Aggregated latency statistics:
- RSS RAM: enabling jemalloc with or without disabling THP decreases RSS RAM consumption by searchd by **5-6%**2)
- enabling jemalloc without disabling THP (se05-2) makes latency worse significantly: avg(response time) - by 23%, 95/99p may be not affected since the differences are within the margin of error (see the control group of se04/04-2 where nothing was changed)
- enabling jemalloc with disabling THP (se03-2) makes:
- avg latency 10% better**
- 95% latency 8% worse
- 99% latency 3.6% better, but this is within the margin of error, so we can’t count this
So in this case we need to decide what we prefer:
- save 5% RAM and have 10% better avg latency, but higher latency spikes for 5% of heavier requests
- or leave all as is: no RAM saving, worse avg latency, but better latency for heavier queries
The results may depend on many things: hardware, workload (amount of data, OPS etc.), but it looks like jemalloc is an interesting thing to try out and can save some RAM/money especially in large installation where saving 5% can mean not buying few new servers.
Let us know if you need our help with applying it on your installation. We can help you with other performance and resources optimizations as well.
P.S. Thanks https://www.sphinxconnector.net/ for the idea to try out jemalloc.