David Kanter (dkanter.delete@this.realworldtech.com) on November 22, 2012 12:12 pm wrote:
> So the only question is whether the value of the additional information
> outweighs the cost of dynamic compilation and GC.

There is a simple rule of thumb for this: bad code runs better in a managed environment. There is no way a a JIT/GC combination can beat well written C/C++ code, but when it comes to bad code the tables are turned. This is because in a managed environment the VM can play any kind of tricks behind your back and modern VMs actually do some amazingly heroic optimizations to make horrible code run fast.

A couple of examples come to my mind, one was a large software that had been made multi-threaded because management decided that it should have been that way, even though the programmers dealing with it had no clue about how to write proper MT code. The result was a monster with a huge amount of locks and critical sections sprinkled all over the code and mostly using thread-safe collections for good measure (even when they where their access was already protected by locks). During execution it spawned 10s of threads but in practice execution was mostly serial. Its biggest drawback however wasn't its mostly serial nature, but the fact that it spent a disproportionate amount of CPU time just taking and releasing useless locks. Fortunately a recent VM was capable of removing almost all of the redundant locks making performance actually OK-ish, something which would have been impossible in a native application.

The other example was another piece of Java code that had an interesting way of dealing with data, this caused enormous amounts of temporary strings to be allocated all the time, to the tune of many GB/s per second. Again such code would have crawled in any non-managed environment (and possibly could have leaked too) but within a VM the GC was coping with it quite well and with a fairly low overhead too (reclaiming short-lived temporary objects is very, very cheap in modern GCs, at high allocation rates it's measurably faster than manual memory management schemes).