<div dir="ltr">On Thu, Aug 14, 2008 at 10:06 PM, Shannon <span dir="ltr"><<a href="mailto:shannon@arc.net.au">shannon@arc.net.au</a>></span> wrote:<br><div class="gmail_quote"><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;">
<div bgcolor="#ffffff" text="#000000">On second thoughts I withdraw these claims. I don't have the statistics
to know one way or the other why "portable threads" are more prevalent
than "share nothing" ones. There may be many reasons but latencies
probably isn't one of them. It could just be fashion or convenience.<br>
</div></blockquote><div> <br>Excuse my rant:<br><br></div></div>Shared-memory threads and locks are a relatively easy model to implement, they *appear* to be a straightforward extension to existing platforms and programming languages, and they don't require much implementation wizardry to provide good performance in simple apps written by really smart people. So they're the natural concurrency extension that everyone adds to their platform first. Then people build really complicated apps that don't scale well and have strange problems with intermittent race conditions and deadlocks and everyone wishes they had a better model, but by then it's too late.<br>
<br>This isn't the right forum to have a big discussion about concurrent programming models. But for the record, here are some of the problems with the shared-memory, threads-and-locks model:<br><br>1) Choosing the scope of locks in time (when you lock and for how long) and space (how much data is covered by each lock) is hard. Mistakes in one direction lead to race conditions where unexpected thread interleavings produce errors, and these are basically impossible to test for. Mistakes in another direction lead to deadlocks which are also very difficult to test for. There are tools that can help detect potential errors but they're no panacea (I did research in this area at IBM).<br>
<br>2) Locks actually scale poorly. It's very hard to achieve high levels of parallelism because you keep hitting lock contention and have to refine lock granularity and then implement various esoteric optimizations to eliminate false sharing etc.<br>
<br>3) It's really hard to do basic stuff well. Maurice Herlihy uses an example of implementing a double-ended queue with concurrent access to both ends; getting it to work correctly for lengths 0, 1, and > 1 is rocket science.<br>
<br>4) Threads and locks just don't support compositional reasoning about programs. To avoid deadlocks and races you have to expose a lot of information about the internals of functions --- what locks they might take, what they might wait for, and what data they might access. For specific domains like OS kernels you can impose rigid rules across the code and get away with it. For more complex apps, especially ones with dynamic extensiblity, that doesn't work well. Especially when not everyone on your team is a genius.<br>
<br>Shared-nothing message passing systems like Workers have problems too and won't be a good fit for some applications, but for the applications that fit, they avoid a lot of problems, and they're really easy to implement.<br>
<br>There are other models, like replacing locks with a notion of atomic code blocks --- gets rid of many problems, but a much harder model to implement efficiently so you don't see it in production systems yet. A related popular idea is adding transactions and transactional memory to the programming language, but again, hard to implement well.<br>
<br>So for the programming model problem, we don't know what the right answer is for all applications, but we do know that threads and locks are the wrong answer.<br><br clear="all">Rob<br>-- <br>"He was pierced for our transgressions, he was crushed for our iniquities; the punishment that brought us peace was upon him, and by his wounds we are healed. We all, like sheep, have gone astray, each of us has turned to his own way; and the LORD has laid on him the iniquity of us all." [Isaiah 53:5-6]<br>
</div>