tldr: Recently, I tried moving an existing solrcloud configuration from a local datacenter to EC2. Performance was roughly 1/10th what I’d expected, until I applied a bunch of linux tweaks.
This should’ve been a straight port: one datacenter server -> one EC2 node. Solr 5.4, Solrcloud, Ubuntu xenial. Nodes were sized in both cases such that the entire index could be cached in memory, and the JVM settings were identical in both places. I applied what should’ve been a comfortable load to the EC2 cluster, and everything exploded. I had to back the rate down to something close to 10% of what I had been getting in the datacenter before latency improved. Looking around, I was interested to note that under load, user-time CPU usage was being shadowed by an almost equal amount of system CPU time. This was not IOWait, but system time. Strace showed a bunch of time being spent in futex and restart_syscall, but I couldn’t see where to go from there. Interestingly, a coworker playing with a ElasticSearch (ES 5.x, so a much more recent release) alternate implementation of the same index was not seeing this high-system-time behavior on EC2, and was getting throughput consistent with our general expectations. Eventually, we came across this: http://www.brendangregg.com/blog/2015-03-03/performance-tuning-linux-instances-on-ec2.html In direct opposition to the author’s intent, (something about taking expired medication) we applied these settings blindly to see what happened. The difference was breathtaking. The system time usage disappeared, and I could apply load at and even a little above my expected rates, well within my latency goals. There are a number of settings involved, and we haven’t isolated for sure which ones made the biggest difference, but my guess at the moment is that it’s the change of clocksource. I think this would be consistent with the observed system time. Note however that using the “tsc” clocksource on EC2 is generally discouraged, because it’s possible to get backwards clock drift. I’m writing this for a few reasons: 1. The performance difference was so crazy I really feel like this should really be broader knowledge. 2. If anyone is aware of anything that changed in Lucene between 5.4 and 6.x that could explain why Elasticsearch wasn’t suffering from this? If it’s the clocksource that’s the issue, there’s an implication that Solr was using tons more system calls like gettimeofday that the EC2 (xen) hypervisor doesn’t allow in userspace. 3. Has anyone run Solr with the “tsc” clocksource, and is aware of any concrete issues?
