MongoDB 2.8: Improving WiredTiger Performance

My (rather popular) first post on this topic explained the benefits of compression (which comes as the default option with the new WiredTiger storage engine) for systems with lesser IO capabilities.  The intent was to first show that the new storage engine saved space on disk and then to show that this could be translated into a gain in terms of performance when reading that data (slowly) off disk.

The first part of that story worked out pretty well, the data was nicely compressed on disk and it was easy to show it in the graph.  The second part of that story did not work out as expected, the graph was a little off from expectations and my initial speculation that it was a non-optimal access pattern didn’t pan out.  In fact, I determined that the slowness I was seeing was independent of IO and was due to how slow the in-memory piece was when using WiredTiger to do a table scan.  Needless to say, I started to talk to engineering about the issue and tried tweaking various options – each one essentially reinforced the original finding.

It was soon obvious that we had a bug that needed to be addressed (one that was still present in the first release candidate 2.8.0-rc0). I gathered the relevant data and opened SERVER-16150 to investigate the problem. Thanks to the ever excellent folks in MongoDB engineering (this one in particular), we soon had the first patched build attempting to address the issue (more, with graphs after the jump).  Before that, for anyone looking to reproduce this testing, I would recommend waiting until SERVER-16150 has been closed and integrated into the next release candidate (2.8.0-rc1), you won’t see the same results from 2.8.0-rc0 (it will instead look like the first set of results).

We may not be completely done in terms of the changes needed to address this fully (please follow

the jira issue if you’d like to see more detail and the ultimate resolution), but the initial patch gave me the improvement I needed to finish off my original post in a far more satisfactory manner.

Figure 1 below now gives me a performance profile which fits with my original reasoning for using compression in a limited IO environment:

Post-patch: time taken to load data set into memory

The two non-compressed engines are similar, with the original engine likely paying the penalty for having a larger overall data size, and this time the graph is far closer to the expected profile.  The main take away is that we can now see the real benefits for trading CPU cycles (via more compression) for IO bandwidth.  The disk being used is not terrible, I can imagine far worse operating conditions than an external USB hard disk with no contention, so it should be possible to make this even more pronounced.  The trend, and comparison, is obvious enough now though, and I think the default choice (snappy) will prove to be a good middle ground for most users.