Nov 16, 2009

NHibernate 2.1 Performance Issues with Flush and LinFu

We recently discovered our nightly processes were taking much longer (50 - 80%) and wondered why. My focus was on the upgrade from 2.0.1.4 straight to 2.1.1. First thing I tried was to change the LinFu Proxy Factory to Castle. This had a surprising boost. For this particular operation some object properties were being accessed 100,000's of times. Redgate Profiler showed that, in aggregate, proxied object access can add up. 3300 objects loaded, with 3 many-to-one objects eagerly loaded, had some some properties accessed 350,000 times. One particular property being accessed took roughly 2 seconds over the life of the process, and some of the others took roughly a second. All adding up to roughly 12 seconds on a 100 second process. Whoa. Keep in mind that's with Redgate's deepest profiling on. Switching to Castle seemed to cut that to 5 seconds. Anyone know what's going on here?

But the big surprise of the day was what happened with flushing from 2.0.1 to 2.1+. A basic flush on this process on 2.0.1 took .42 seconds, but on 2.1.1 it took .79 seconds. Really? What the hell happened in between those versions? Just to be sure I checked 2.1 and 2.0.1, both were the same. I wonder if abstracting dynamic proxying from NH had some unintended consequences.

Needless to say we are back on 2.0.1 for the time being. This process that was taking 1 min and 11 seconds now takes 47 seconds. Why? A couple of reasons: first the process calls embedded business logic many times that ends up doing "safety" flushing when it's a mostly readonly session.  The next refactor is to fix that problem, but in all the flush increase difference accounted for 10 - 15 seconds.  Second, was the proxies which accounted for the remainder of the difference. I didn't do anymore benchmarks on 2.0.1 vs Castle on 2.1.1, but I'm sure there are differences there from all the testing that was done. In either case there's not anything in 2.1+ that's worth these performance tradeoffs we experienced. NH.Linq seemed like it would have been a good move, but it's not yet ready for production use. So for now I'll take the perf benefits of 2.0.1 for the minor upgrades of 2.1+.

2 comments:

  1. Nope, just Castle and LinFu. Haven't looked at the Spring stuff too much. Might be worth looking at sometime.

    ReplyDelete
  2. Get newer Castle Dynamic Proxy (v2.2) (you may want to download Castle ActiveRecord to obtain Castle.ByteCode, sinec NHibernate does not provide one).

    New version is faster when you generate many different proxy types (actual usage of proxies should have around the same performance).

    ReplyDelete