Show Posts
1
BitShares PTS / Algorithm efficiency with memory usage
« on: November 17, 2013, 09:14:59 am »
Hello,
I did a search of the forum but unfortunately wasn't coming up with much in way of a question this side of the fence.
(This is mostly a question specifically for the algo designer / developers)
I'm interested to know if there is actually a theoretical (eg, based on the algorithm/code specifically) of running with higher memory usage (eg, default is 512, but while some are working toward lower memory usage there is a mod by Tydus to permit 1024/2048/4096 MB RAM per thread.
I haven't seen a terrible difference in collisions/min when testing between these (in fact running with 2048MB appears to be 1-5% more col/min than 4096MB (Which I assume is a memory bandwidth/latency issue).
Generally speaking however what I'd like to know is - all things being equal is more memory actually better per thread where possible? Is there a larger hash table stored in memory to use as a reference? (Meaning that it may be more likely to find a meaningful collision?) or is the thread only going to operate within a small amount of that hash space?
At the moment I'm trying to do some tests myself (Which is as simple as running the program at different settings at ~24H at a time to see if one gets more shares and/or consistently runs higher col/min) but as far as I can tell those statistics would be completely arbitrary as it could be complete fluke whether or not a collision is valid or if there's a pattern to it.
Overall if I don't make sense feel free to ask (or yell at me if it's been explained already somewhere else, I unfortunately wasn't sure what keywords to search for beyond a couple that didn't seem to get me the answer I was looking for).
Thanks,
I did a search of the forum but unfortunately wasn't coming up with much in way of a question this side of the fence.
(This is mostly a question specifically for the algo designer / developers)
I'm interested to know if there is actually a theoretical (eg, based on the algorithm/code specifically) of running with higher memory usage (eg, default is 512, but while some are working toward lower memory usage there is a mod by Tydus to permit 1024/2048/4096 MB RAM per thread.
I haven't seen a terrible difference in collisions/min when testing between these (in fact running with 2048MB appears to be 1-5% more col/min than 4096MB (Which I assume is a memory bandwidth/latency issue).
Generally speaking however what I'd like to know is - all things being equal is more memory actually better per thread where possible? Is there a larger hash table stored in memory to use as a reference? (Meaning that it may be more likely to find a meaningful collision?) or is the thread only going to operate within a small amount of that hash space?
At the moment I'm trying to do some tests myself (Which is as simple as running the program at different settings at ~24H at a time to see if one gets more shares and/or consistently runs higher col/min) but as far as I can tell those statistics would be completely arbitrary as it could be complete fluke whether or not a collision is valid or if there's a pattern to it.
Overall if I don't make sense feel free to ask (or yell at me if it's been explained already somewhere else, I unfortunately wasn't sure what keywords to search for beyond a couple that didn't seem to get me the answer I was looking for).
Thanks,
