On 22 March 2011 02:00, Jesper Louis Andersen wrote:

On Tue, Mar 22, 2011 at 00:59, David MacIver wrote:

...

It's for rank aggregation - taking a bunch of partial rankings of some items from users and turning them into an overall ranking (aka "That thing that Hammer Principle does").

Two questions immediately begs themselves:

* Can we go parallel? :P

Maybe. A lot of this is inherently sequential. Some bits are parallelisable, but my initial attempts at exploiting that made very little performance difference. I'd rather exhaust what I can from single-core performance first.

* What does +RTS -s -RTS say? Specifically, what is the current productivity?

./rank +RTS -s 3,466,696,368 bytes allocated in the heap 212,888,240 bytes copied during GC 51,949,568 bytes maximum residency (10 sample(s)) 5,477,016 bytes maximum slop 105 MB total memory in use (0 MB lost due to fragmentation) Generation 0: 6546 collections, 0 parallel, 0.93s, 0.93s elapsed Generation 1: 10 collections, 0 parallel, 0.32s, 0.32s elapsed INIT time 0.00s ( 0.00s elapsed) MUT time 7.11s ( 7.12s elapsed) GC time 1.25s ( 1.25s elapsed) EXIT time 0.00s ( 0.00s elapsed) Total time 8.37s ( 8.37s elapsed) %GC time 15.0% (15.0% elapsed) Alloc rate 487,319,292 bytes per MUT second Productivity 85.0% of total user, 85.0% of total elapsed So if I'm reading this right, my hypothesis that allocation was most of the cost seems to be wrong? I don't know how much of that MUT time is allocation, but I'd expect it to be < GC time.

Do we get an improvement with +RTS -A2m -H128m -RTS ? (Force the heap to be somewhat up there from day one, perhaps try -H256m.

This seems to consistently give about a 0.4s improvement, which isn't nothing but isn't a particularly interesting chunck of 8s (actually it's 8.4s -> 8s). Setting it to 256M doesn't make any difference.

On Tue, Mar 22, 2011 at 1:11 AM, David MacIver wrote:

On 22 March 2011 02:00, Jesper Louis Andersen wrote:

...

On Tue, Mar 22, 2011 at 00:59, David MacIver wrote:

...

It's for rank aggregation - taking a bunch of partial rankings of some items from users and turning them into an overall ranking (aka "That thing that Hammer Principle does").

Two questions immediately begs themselves:

* Can we go parallel? :P

Maybe. A lot of this is inherently sequential. Some bits are parallelisable, but my initial attempts at exploiting that made very little performance difference. I'd rather exhaust what I can from single-core performance first.

...

* What does +RTS -s -RTS say? Specifically, what is the current productivity?

./rank +RTS -s 3,466,696,368 bytes allocated in the heap 212,888,240 bytes copied during GC 51,949,568 bytes maximum residency (10 sample(s)) 5,477,016 bytes maximum slop 105 MB total memory in use (0 MB lost due to fragmentation)

Generation 0: 6546 collections, 0 parallel, 0.93s, 0.93s elapsed Generation 1: 10 collections, 0 parallel, 0.32s, 0.32s elapsed

INIT time 0.00s ( 0.00s elapsed) MUT time 7.11s ( 7.12s elapsed) GC time 1.25s ( 1.25s elapsed) EXIT time 0.00s ( 0.00s elapsed) Total time 8.37s ( 8.37s elapsed)

%GC time 15.0% (15.0% elapsed)

Alloc rate 487,319,292 bytes per MUT second

Productivity 85.0% of total user, 85.0% of total elapsed

So if I'm reading this right, my hypothesis that allocation was most of the cost seems to be wrong? I don't know how much of that MUT time is allocation, but I'd expect it to be < GC time.

...

Do we get an improvement with +RTS -A2m -H128m -RTS ? (Force the heap to be somewhat up there from day one, perhaps try -H256m.

This seems to consistently give about a 0.4s improvement, which isn't nothing but isn't a particularly interesting chunck of 8s (actually it's 8.4s -> 8s). Setting it to 256M doesn't make any difference.

You should use criterion to make sure that your assessments of the performance difference are grounded in statistically robust reasoning :) Progression may also help. GHC 7 has a new rts option, +RTS -H -RTS that gives you a good high default for these memory numbers. You might give it ago, but I doubt it's going to help much here. One thing I noticed was that you calculate the length of elts more than once inside aggregate. It's probably not going to account for a big boost, but I'd recommend doing it only once. Something like: let lenelts = length elts Then use lenelts everywhere. Also, it seems like you use Data.Set to sort and uniquify the input then you throw away the Set, that's potentially costly. There are probably better types for aggregate than [[a]] -> [a]. You use Arrays, but it's possible that a different vector library such as Data.Vector gives better performance here. Or, perhaps you want that instead of lists everywhere. For example, you build an index and a reverse index. It seems like with an array or vector you could eliminate at least one index and have O(1) lookup time instead of Data.Map's O(log n) lookup time. I hope that helps, Jason