On Tue, 2004-07-27 at 15:32, Marc Kleine-Budde wrote: > On Tue, Jul 27, 2004 at 03:22:47PM +0200, Marc Kleine-Budde wrote: > > > If you want to compare fusion with LXRT, use the testsuite/latency test, > > > it has been crafted exactely for that. On my 1Ghz laptop, your "dd" > > > test gives the following, which is rather correct: > > > > min = 4130 ns, max = 46003 ns, avg = 4908 ns, overrun = 0 > > > min = 4130 ns, max = 46003 ns, avg = 4900 ns, overrun = 0 > > > min = 4130 ns, max = 46003 ns, avg = 4909 ns, overrun = 0 > > > min = 4130 ns, max = 46003 ns, avg = 4904 ns, overrun = 0 > > > min = 4130 ns, max = 46003 ns, avg = 4910 ns, overrun = 0 > > > min = 4130 ns, max = 46003 ns, avg = 4905 ns, overrun = 0 > > > min = 4130 ns, max = 46003 ns, avg = 4912 ns, overrun = 0 > > > min = 4130 ns, max = 46003 ns, avg = 4910 ns, overrun = 0 > > here: > > using fusion-0.3-2.6-i386/testsuite/latency > > > min = 5185 ns, max = 79593 ns, avg = 10617 ns, overrun = 0 > > min = 5185 ns, max = 79593 ns, avg = 10797 ns, overrun = 0 > > min = 5185 ns, max = 79593 ns, avg = 11028 ns, overrun = 0 > > min = 5185 ns, max = 79593 ns, avg = 10979 ns, overrun = 0 > > min = 5185 ns, max = 79593 ns, avg = 9906 ns, overrun = 0 > > min = 5185 ns, max = 79593 ns, avg = 9144 ns, overrun = 0 > > min = 5185 ns, max = 79593 ns, avg = 11602 ns, overrun = 0 > > > But with the original lxrt, we've still a third (~20us) max. > okay only the half... > > using rtai-3.1-2.6-i386/testsuite/user/latency > > 2004/07/27 15:28:06 min: 111 max: 35884 average: 4396 > 2004/07/27 15:28:07 min: 111 max: 35884 average: 4316 > 2004/07/27 15:28:08 min: 111 max: 35884 average: 4916 > 2004/07/27 15:28:09 min: 111 max: 35884 average: 4938 >
What would be interesting is having a distribution of the peaks for fusion/latency. I wonder if this is a one shot problem at startup, or a continuous jitter spreaded all over the test. In any case, there is at least a ~4 us inherent jitter as shown by min values which is not correctly taken in account by the calibration code for either/both timer and scheduling. But this obviously does not explain all. At the very least, did you experiment any unbounded jitter while testing? > marc -- Philippe.
