| Author | Message | ||
     Pánczél, Levente Username: panczel_levente Registered: N/A |
( Sorry, if I double post, I wanted to write about this for quite some time now, but if I did write then I simply cannot find the thread. ) On a my quad machine SHA-256 calculation seems to have a top speed of 5500 MB/min. (Excellent speed!) But when I start multiple instances that also use SHA-256 (e.g. imaging or verification) the the sum-speed rapidly degrades. Until now I thought windows would mess something with I/O, but now I've tested it on sparse files (full-zero), and they show the same sympthoms (although I verified that XWF can access such files with at least 30000 MB/s in any number of instances). E.g. if I run 2 instances hashing the same sparse file, the sum-speed drops below 3500 MB/s, and CPU utilization is about 60% (reasonable for 2 hash threads occupying 50% plus the rest). So two instances, twice CPU usage, planty of I/O capacity free (tested: when waiting for I/O, the threads lose CPU utilization, e.g. hasher threads do not use 25% [full core] anymore); but a speed drop of around 60%. I also found that it is irrelevant which copies I run: 2 totally distinct copies block each other just the same. Also checked whether they load the same hash.dll, but they don't. It rather seems that hashing is using some other shared resorce, I cannot think of other explanation for the mutual exclusion. I know the above wasn't very detailed I just wanted to outline the concept, and this should be easy to reproduce. If you have troubles with testing this behavior then I will of course write detailed scenarios. It would be very handy if instances would not affect the speed of each other (as long as the resources used permit this). We often use the approach of imaging with one instance while the other is verifying the image of the previous disk. Now these two processes block each other so much that reading the source disk stops and stalls for 1 second in every two seconds (while hashing is paused, imaging runs normally). So when having to work on one computer (even if it's a powerful one) we end up with the write blocker being half the time idle. (I can confirm that when the new imaging is started without hash calculation then the two instances seem not to interfere. Preliminary test on other harware/OS combinations gave analogous results.) | ||
| Stefan Fleischmann Username: admin Registered: 1-2001 |
I cannot reproduce such a problem here, also not this problem. |