Cache effects on benchmark

Jump to navigation Jump to search

Cache effects on benchmark

I was thinking that running *just* the benchmark vs a single tree at a time would result in the KD-Tree code and data being cached quite a bit better than would be realistic for real-life situations. Perhaps it would make more sense to run all of the trees at the same time, giving each tree the new search/data one tree after the next. This would simulate the cache thrashing between turns that happens when you are running multiple robots at the same time, because the trees would be competing with each other for cache space.

Thoughts?

The reason I ask is that I designed/wrote a tree to deal with the cache problem. It outperforms Rednaxela Gen2 on large (2mil points, 12 dim) random datasets by ~2X, but ties on smaller (30k point, 12 dim) datasets. I think the fact that for the small dataset the entire thing is in cache might be causing the difference.

Skilgannon (talk)21:03, 16 July 2013

Maybe using a reference bot would make benchmarking more meaninful?

Pick one bot, put each tree inside it, one at a time, and run it against a 1v1 test bed.

MN (talk)01:37, 17 July 2013

For the best comparison, absolutely. However, it might be difficult to set up the battles so that every one is the same, particularly if the trees are non-deterministic due to things like points being equal. Also, it adds a lot of overhead which would make testing very slow.

Skilgannon (talk)11:56, 17 July 2013

It would make testing slow, but the overhead is what will make benchmarks meaningful. Remove the overhead and you also remove cache thrashing.

Running a test bed much like we run challenges would do. Instead of making every battle the same, run multiple random battles and measure average run time.

MN (talk)13:06, 17 July 2013

What I worry is that many different trees would have to be tested in exactly the same way. Unlike scores, times are different on different computers. Perhaps if we put them all in parallel in the same robot, then see how much time they take?

Skilgannon (talk)15:57, 17 July 2013
 
 
 

Running them all at the same time could make sense, but I would suggest being careful if you do that, because the order that they get run in may matter. Even when running them one at a time in sequence, I've recall noticing that the order in which they are run could very slightly impact the apparent performance, I suspect due to caching, JIT, and/or garbage collection characteristics. It's been a while, but IIRC the System.gc() call I have in there between running different trees was to lessen that effect. It may make sense to add some form of randomization to sequence they're run in.

Cache performance is one of those things that's tricky with Robocode, because your robot is also sharing the CPU with another bot which could be doing who knows that with it's memory accesses. For that reason I wouldn't trust optimizations for better caching behavior to necessarily pan out in practice with bots. I may be wrong about that though.

One could put it in a reference bot yeah, though the tests would be far slower and less consistent, thus requiring a much greater number of test iterations to have a reliable result.

Oh, and in any case, nice job with making a tree that much faster with the large datasets :)

Rednaxela (talk)05:13, 17 July 2013

I wrote a quick benchmark and threw it in my main method. Some tweaks and improvements (putting the distance functions in methods, using a binary search for the results array) have given me big improvements at lower tree sizes:

Config:
No JIT Warmup
Tested on random data.
Training and testing points shared across iterations.
Searches interleaved.
Num points:     20000
Num searches:   200
Dimensions:     12
Num Neighbours: 40

Accuracy: 100%
Iteration:      1/3
This tree add avg:  3092 ns
Reds tree add avg:  2216 ns
This tree knn avg:  809965 ns
Reds tree knn avg:  1380366 ns
This tree knn max:  10844097 ns
Reds tree knn max:  11005183 ns

Accuracy: 100%
Iteration:      2/3
This tree add avg:  1259 ns
Reds tree add avg:  846 ns
This tree knn avg:  643037 ns
Reds tree knn avg:  1119268 ns
This tree knn max:  979013 ns
Reds tree knn max:  1787566 ns

Accuracy: 100%
Iteration:      3/3
This tree add avg:  1146 ns
Reds tree add avg:  800 ns
This tree knn avg:  641163 ns
Reds tree knn avg:  1099657 ns
This tree knn max:  1318587 ns
Reds tree knn max:  1782212 ns

Note, I hacked the RedGen2 tree to not check for NaN in the distance functions so that they are on equal footing.

Skilgannon (talk)12:03, 17 July 2013

Ahh nice. Out of curiosity, any reason you're comparing against my 2nd gen tree? My 3rd gen tree was a bit faster at least in the tests I did.

Rednaxela (talk)14:00, 17 July 2013

I don't have it installed yet, and I thought I'll do a proper bench when I add my tree to the whole bench framework. Do you still have the code that does those nice charts tracking the trees through time? Is that what's in the KNN.jar? Yes it is! Great.

Skilgannon (talk)15:40, 17 July 2013

Ah, looks like KNN.jar that's uploaded on the wiki isn't quite 100% up-to-date. A couple days after the KNN.jar uploaded to the wiki was last updated I made a few minor changes. See here for the latest source [1]

Rednaxela (talk)17:46, 17 July 2013

You got that just in time, your latest code is now in!

So on the vanilla Diamond-gun data benchmark it seems to be pretty much tied with your Gen2 and Gen3, changing places from run to run. This is what I saw mostly though, your Gen2 slightly above Gen3, and my cache-hit tree in the middle:

BEST RESULT: 
 - #1 Rednaxela's kd-tree (2nd gen) [0.0317]
 - #2 Skilgannon's Cache-hit KDTree [0.0340]
 - #3 Rednaxela's kd-tree (3rd gen) [0.0356]
 - #4 Voidious' Linear search [0.4175]

So, I wrote some code which instantiates, does operations on, then destroys an array of 1k doubles and 1k ints, and makes it a dependency of a later println. This was the result:

RESULT << k-nearest neighbours search with Voidious' Linear search >>
: Average searching time       = 0.4198 miliseconds
: Average worst searching time = 14.7468 miliseconds
: Average adding time          = 1.6292 microseconds
: Accuracy                     = 100.0%

RESULT << k-nearest neighbours search with Rednaxela's kd-tree (2nd gen) >>
: Average searching time       = 0.0333 miliseconds
: Average worst searching time = 1.7135 miliseconds
: Average adding time          = 1.7662 microseconds
: Accuracy                     = 100.0%

RESULT << k-nearest neighbours search with Skilgannon's Cache-hit KDTree >>
: Average searching time       = 0.0312 miliseconds
: Average worst searching time = 1.3695 miliseconds
: Average adding time          = 2.7642 microseconds
: Accuracy                     = 100.0%

RESULT << k-nearest neighbours search with Rednaxela's kd-tree (3rd gen) >>
: Average searching time       = 0.0317 miliseconds
: Average worst searching time = 1.2998 miliseconds
: Average adding time          = 2.1558 microseconds
: Accuracy                     = 100.0%


BEST RESULT: 
 - #1 Skilgannon's Cache-hit KDTree [0.0312]
 - #2 Rednaxela's kd-tree (3rd gen) [0.0317]
 - #3 Rednaxela's kd-tree (2nd gen) [0.0333]
 - #4 Voidious' Linear search [0.4198]

It now comes first every single time I run. So it seems my cache-hit design has paid off (slightly) =) Cache kdtree vs RedG2G3.png

Skilgannon (talk)19:07, 17 July 2013

Ahhh neat. Hmm... having had a quick look at your code earlier today, I'm feeling a little tempted to add some cache-behavior optimizations to my tree... oh wow... just realized it's been 3 years since I touched it.

Rednaxela (talk)19:39, 17 July 2013

Feel free to port it to Lua while you're at it... :-) Just kidding, but fyi I might sometime. The first time I tried to write a BerryBots gun and realized I didn't have a kd-tree to work with was sobering. :-)

Voidious (talk)19:41, 17 July 2013