Messages in this thread | | | Date | Sat, 01 Feb 2003 12:23:19 +1100 | From | Nick Piggin <> | Subject | Re: [BENCHMARK] 2.5.59-mm7 with contest |
| |
Con Kolivas wrote:
>On Saturday 01 Feb 2003 11:55 am, Nick Piggin wrote: > >>Con Kolivas wrote: >> >>>On Saturday 01 Feb 2003 11:37 am, Nick Piggin wrote: >>> >>>>Con Kolivas wrote: >>>> >>>>>Seems the fix for "reads starves everything" works. Affected the tar >>>>>loads too? >>>>> >>>>Yes, at the cost of throughput, however for now it is probably >>>>the best way to go. Hopefully anticipatory scheduling will provide >>>>as good or better kernel compile times and better throughput. >>>> >>>>Con, tell me, are "Loads" normalised to the time they run for? >>>>Is it possible to get a finer grain result for the load tests? >>>> >>>No, the load is the absolute number of times the load successfully >>>completed. We battled with the code for a while to see if there were ways >>>to get more accurate load numbers but if you write a 256Mb file you can >>>only tell if it completes the write or not; not how much has been written >>>when you stop the write. Same goes with read etc. The load rate is a more >>>meaningful number but we haven't gotten around to implementing that in >>>the result presentation. >>> >>I don't know how the contest code works, but if you split that into >>a number of smaller writes it should work? >> > >Yes it would but the load effect is significantly diminished. By writing a >file the size==physical ram the load effect is substantial. > Oh yes of course, but I meant just break up the writing of that big file into smaller write(2)s.
> > >>>Load rate would be: >>> >>>loads / ( load_compile_time - no_load_compile_time ) >>> >>I think loads / time_load_ran_for should be ok (ie, give you loads per time >>interval). This would be more useful if your loads were getting more >>efficient >>or less because it is possible that an improvement would lower compile time >>_and_ loads, but overall the loads were getting done quicker. >> > >I found the following is how loads occur almost always: >noload time: 60 >load time kernal a: 80, loads 20 >load time kernel b: 100, loads 40 >load time kernel c: 90, loads 30 > >and loads/total time wouldnt show this effect as kernel c would appear to have >a better load rate > Kernel a would have a rate of .25 l/s, b: .4 l/s, c: .33~ l/s so I b would be better.
> > >if there was >load time kernel d: 80, loads 40 > >that would be more significant no? > It would, yes... but it would measure .5 loads per second done.
The noload time is basically constant anyway so I don't think it would add much value if it were incorporated into the results, but would make the metric harder to follow than simple "loads per second".
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |