Naive parallelism with HLVM

The latest OCaml Journal article High-performance parallel programming with HLVM (23rd January 2010) described a simple but effective parallelization of the HLVM implementation of the ray tracer. Comparing with similarly naive parallelizations of the fastest serial implementations in C++ and Haskell, we obtained the following results:

These results exhibit several interesting characteristics:

  • Even the naive parallelization in C++ is significantly faster than HLVM and Haskell.
  • C++ and HLVM scale well, with performance improving as more cores are used.
  • Despite having serial performance competitive with HLVM, the naively-parallelized Haskell scales poorly. In particular, Haskell failed to obtain a competitive speedup with up to 5 cores and its performance even degraded significantly beyond 5 cores, running 4.4× slower than C++ on 7 cores.

The efficiency of the parallelization can be quantified as the speed of the parallel version on multiple cores relative to its speed on a single core:

These results show that HLVM obtained the largest speedup: 6.3× faster on 8 cores. C++ obtained the next largest speedup: 5.2x faster on 8 cores. Haskell obtained the worst speedup: only 2.9× faster on 8 cores.

More sophisticated parallelizations could obtain better performance still. Currently, the parallel for loop over the rows of the image causes each thread to compete for an iteration and results in poor locality: if a thread computes one row then it is unlikely to compute the next. Work-stealing queues and recursive subdivision of the problem space would greatly improve locality and, therefore, performance. Also, the initial construction of the scene tree has some opportunity for parallelization and further performance improvements.

Furthermore, there is plenty of room to optimize HLVM itself and some simple hacks can quantify the possibilities. Disabling bounds checking provides a substantial 20% performance improvement. Disabling the shadow stack (and, therefore, disabling GC) provides another 25% performance improvement. With these two changes, HLVM is only 24% slower than C++. In particular, HLVM's current design makes heavy reuse of its own generic routines even in performance critical sections such as manipulating the shadow stack. Optimizing these routines by hand could leverage some of this potential. For example, by removing bounds checks from manipulations of the shadow stack.

Comments

Post a Comment

Popular posts from this blog

Bjarne Stroustrup is catching up

Bjarne Stroustrup, creator of the C++ programming language, once famously said "There are only two kinds of languages: the ones people complain about and the ones nobody uses". Interestingly, Bjarne has gone on the defensive in his recent lectures, completely changed his tune and is catching up with the conclusions that most former-C++ developers have arrived at. In a recent lecture Bjarne made many eyebrow-raising assertions. He is happy that people are no longer talking about C++ because that means it has succeeded. In reality, C++ demand in the job market has been in freefall for years and few new software projects are choosing it. He attacked computer scientists for copying data and said that "even babies don't do that", a very strange statement to make in a technical presentation. He also implied that other languages deep copy 10,000x10,000 matrices and claimed that a shared_ptr is "like an old fashioned garbage collector except it gets resource relea...
Read more

Does reference counting really use less memory than tracing garbage collection? Mathematica vs Swift vs OCaml vs F# on .NET and Mono

Image
Our previous post caused some controversy by questioning the validity of some commonly-held beliefs. Specifically, the beliefs that reference counting (RC) always requires less memory than tracing garbage collection (GC) and that tracing GCs require 3-4x more memory in order to perform adequately. We presented a benchmark written in Swift and OCaml and noted that the RC'd Swift implementation ran over 5x slower and required over 3x more memory than the tracing GC'd OCaml implementation. This observation disproves these beliefs in their strong form and even brings into question whether there is even any validity in the weak forms of those beliefs. After all, we have never seen any empirical evidence to support these beliefs in any form. We received a lot of criticism for that post. The vast majority of the criticism was not constructive but two valid points did arise. Firstly, although our result was anomalous it would be more compelling to see the benchmark repeated across a w...
Read more

Does reference counting really use less memory than tracing garbage collection? Swift vs OCaml

Image
The way software developers cling to folklore and avoid proper experimental testing of hypotheses really bothers me. I think perhaps the worst branch of computer science still riddled with myths and legends is memory management. Despite over half a century of research on garbage collection showing that reference counting is inferior to tracing garbage collections algorithms (even when almost all GC research restricts consideration to Java when much better algorithms have been known for years) there are still many people who claim otherwise. C++ developers obviously still believe that reference counted smart pointers are superior to tracing garbage collection but now other people are doing it too. People who should know better. This situation recently reared its ugly head again when Apple released a promising new programming language called Swift that shuns tracing garbage collection in favor of reference counting. Now, there are logical reasons for Apple to have chosen reference count...
Read more