undefined | Better HN

0 pointsvlovich1239mo ago0 comments

We actually do have the abstractions but the problem is that the vast majority of for loops don’t benefit - you need to have so much work that the overhead of coordinating the threads is sufficient. Additionally, you’ve got all sorts of secondary effects like cache write contention that will fight any win you try to extract out of for loops parallelism. What we’ve been learning for a long time as an industry is that you benefit most from task level parallelism with minimal to no synchronization.

0 comments

dkarl9mo ago

Granted this probably isn't the parallel application that the other poster was envisioning, but it can be extremely useful when a computation depends on a large number of I/O-bound tasks that may fail, like when you are servicing a request with a high fan-out to other services, and you need to respond in a fixed time with the best information you have.

For example, if you need to respond to a request in 100ms and it depends on 100 service calls, you can make 100 calls with a 80ms timeout; get 90 quick responses, including two transient errors, and immediately retry the errors; get eight more successful responses and two timeouts; and then send the response within the SLA using the 98 responses you received.

vlovich123OP9mo ago

That doesn't require parallelism, just concurrency. But yes, you'd use a similar task-local map/reduce construct to express doing a bunch of concurrent I/O in parallel (spawning each I/O on a separate thread would be counter-productive & a hack to enable not adding an event loop / async I/O).

fifilura9mo ago

Let the engine or compiler decide if it is small enought to run on one core.

bee_rider9mo ago

I’m only familiar with Fortran and my openMP is a little rusty. But, I think there are different pragmas for vectorization or threading. So, you have to tell it to do one or the other (is that wrong). Instead of expressing “well we don’t have dependencies here so do what you will.”

vlovich123OP9mo ago

That would be great if the engine or compiler had that kind of capability but building that requires solving the halting problem.

Even if you try to do it with heuristics, go ask Itanium how that worked out for them and they tried a much simpler problem than what you’re proposing.

fifilura9mo ago

I am not familiar with the Itanium story and I don't know who to ask.

But it seems to me like this would be a safe space to experiment. With heuristics and pragmas as a fallback. Because with the right approach solutions would mostly be better than not doing anything.

And you could do it in runtime when you know the size of the input.

And what about applying the logic to places where you can see that the loop will end?

I believe query planners in for example Trino/BigQuery do this already?

1 more reply

j / k navigate · click thread line to collapse

0 comments

dkarl9mo ago

vlovich123OP9mo ago

fifilura9mo ago

Let the engine or compiler decide if it is small enought to run on one core.

bee_rider9mo ago

vlovich123OP9mo ago

That would be great if the engine or compiler had that kind of capability but building that requires solving the halting problem.

Even if you try to do it with heuristics, go ask Itanium how that worked out for them and they tried a much simpler problem than what you’re proposing.

fifilura9mo ago

I am not familiar with the Itanium story and I don't know who to ask.

But it seems to me like this would be a safe space to experiment. With heuristics and pragmas as a fallback. Because with the right approach solutions would mostly be better than not doing anything.

And you could do it in runtime when you know the size of the input.

And what about applying the logic to places where you can see that the loop will end?

I believe query planners in for example Trino/BigQuery do this already?

1 more reply

j / k navigate · click thread line to collapse