After my quick last post I decided to look more into what I could do to save the one year of work I pushed into Ruby-Elf. After fighting a bit to understand how the new String class from Ruby 1.9 worked, I could get the testsuite to pass on Ruby 1.9, and cowstats to also report the correct data.
Unfortunately, this broke down Ruby 1.8 support, as far as I can see because IO#readpartial does not work that well on 1.8 even for on-disk files; similarly happens to JRuby.
After getting Ruby 1.9 to work the obvious next task was to make cowstats work with multiple threading. The final idea was to have a -j parameter akin to make’s but for now I only wanted to create one thread per file to scan. In theory, given native threading, all the threads would be executing at once, scheduled by the system schedule, allowing to saturate the 8 cores, reaching 800% of CPU time as a theoretical maximum.
Unfortunately reality soon kicked in, and the ruby19 process limited itself to 100%, which means a single core out of eight, which also means no parallel scan. A quick glance through the sources shows that while YARV (the Ruby 1.9 VM) lists three possible methods to achieve mutlithreading, only one is currently implemented, the second one. The first method is the old one, green threading, which basically means simulated threads, as the code never executes in parallel but uses an event-loop-like construct to switch the execution between different “threads”. The second method makes use of a giant lock, which in this case is called Giant VM Lock (GVL), and is called GIL (Giant Interpreter Lock) in Python, where the threads are scheduled by the operating system, which allows for more fair scheduling among execution threads, but still allows just one thread per VM to be executed in parallel. The third method is the one I was hoping for and allows for multiple threads to be executed at the same time on different cores on the same virtual machine; instead of having a single lock on the whole VM, the locks are sparse around the code to just lock the needed resources for each thread.
I also checked this out on JRuby, to compare; unfortunately JRuby in portage cannot handle the code as I changed it to work with Ruby 1.9, so I have been unable to actually benchmark a working run of cowstats with it; but I could see that the CPU used by JRuby spiked at 250%, which means it at least is able to execute the threads quite independently; which proves that Ruby can be parallelised up to that point just fine.
So what is the fuss about Ruby 1.9 new native threading support if multiple threads cannot be executed in parallel? Well it still allows for a single process to spawn multiple VMs and execute parallel threads on them, isolated one from the other. Which happens to be useful for Ruby on Rails web applications. If you think well about it, the extra complexity added to deal with binary files is also to address some interesting problems that come up in environment where multiple encodings can often be used, which is, web applications. Similarly the JRuby approach, which is very fast once the JVM is loaded, works fine for applications where you start up once and then proceed to elaborate for a long time, which again fits web application and little more.
I’m afraid to say that what we’re going to see in the next and not-so-next future is for Ruby to lose the general-purpose support and just focus more and more on the web application side of the fence. Which is sad since I really cannot think of anything else I would like to rewrite my tools in, beside, maybe, C# (if it could be compiled in ELF — I should try Vala for that). I feel like my favourite general-purpose language is slipping away, and I should stop worrying and working on that.