|Just another Perl shrine|
Optimizing with Caching vs. Parallelizing (MCE::Map)by 1nickt (Abbot)
|on Apr 05, 2020 at 15:17 UTC||Need Help??|
Mon cher ami Laurent_R recently blogged about his solution to the "extra credit" problem in the Perl Weekly Challenge #54. He showed a solution using memoizing, or caching, to reduce the number of repeated calculations made by a program.
I wondered about the strategy. Obviously calculating the sequences for numbers up to 1,000,000 without some optimization would be painfully or maybe unworkably slow. But the task looks computation-intensive, so I wanted to see whether more cycles would be more beneficial than caching.
Here is the solution presented by Laurent:
This runs on my system pretty quickly:
Next I ran the following version using mce_map_s from MCE::Map. mce_map_s is an implementation of the parallelized map functionality provided by MCE::Map, optimized for sequences. Each worker is handed only the beginning and end of the chunk of the sequence it will process, and workers communicate amongst themselves to keep track of the overall task. When using mce_map_s, pass only the beginning and end of the sequence to process (also, optionally, the step interval and format).
This program, with no caching, runs on my system about five times faster (I have a total of 12 cores):
Notably, reducing the number of workers to just two still ran the program in less than half the time than Laurent's single-process memoized version. Even running with one process, with no cache, was faster. This is no doubt due to the fact MCE uses chunking by default. Even with one worker the list of one million numbers was split by MCE into chunks of 8,000.
Next, I implemented Laurent's cache strategy, but using MCE::Shared::Hash. I wasn't really surprised that the program then ran much slower than either previous version. The reason, of course, is that this task pretty much only makes use of the CPU, so while throwing more cycles at it it a huge boost, sharing data among the workers - precisely because the task is almost 100% CPU-bound - only slows them down. Modern CPUs are very fast at crunching numbers.
I was curious about how busy the cache was in this case, so I wrapped the calls to assign to and read from the hash in Laurent's program in a sub so I could count them. The wrappers look like:
That's a lot of back and forth.
So the moral of the story is that while caching is often useful when you are going to make the same calculations over and over, sometimes the cost of the caching exceeds the cost of just making the calculations repeatedly.
Hope this is of interest!