| Message ID | 20201218141935.24151-5-dwagner@suse.de |
|---|---|
| State | Superseded |
| Headers | show |
| Series | libnuma cleanups for cyclictest | expand |
On Fri, 18 Dec 2020, Daniel Wagner wrote: > Let --affinity without an argument behave in the same way as > --smp. Run a thread on each CPU. This makes cyclictest behave as the > rest of the rt-tests when --affinity is used. > > Reviewed-by: Daniel Wagner <dwagner@suse.de> > --- > src/cyclictest/cyclictest.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/src/cyclictest/cyclictest.c b/src/cyclictest/cyclictest.c > index 0e6519125c2f..b9b0eb3575e3 100644 > --- a/src/cyclictest/cyclictest.c > +++ b/src/cyclictest/cyclictest.c > @@ -1024,6 +1024,8 @@ static void process_options(int argc, char *argv[], int max_cpus) > argv[optind][0] == '0' || > argv[optind][0] == '!')) { > parse_cpumask(argv[optind], max_cpus, &affinity_mask); > + } else { > + num_threads = -1; > } > > if (verbose) > -- > 2.29.2 > > Well, --smp historically combined -a -t and threads at the same priority You could argue that it is reasonable for -a to automatically imply -t but I have had debates with people about this and we settled on -a just specifies the affinity, and the default number of threads is one unless you use -t. I'm not sure what you mean by this makes cyclictest behave the way the rest of rt-tests does, the rest of rt-tests should match what cyclitest does. That said, I did some quick runs of signaltest and -a seems broken, sigh.
On Tue, Jan 26, 2021 at 12:55:33AM -0500, John Kacur wrote: > Well, --smp historically combined -a -t and threads at the same priority > You could argue that it is reasonable for -a to automatically imply -t > but I have had debates with people about this and we settled on -a just > specifies the affinity, and the default number of threads is one unless > you use -t. My thinking is if you set an affinity mask, you want also to assign a thread to the CPUs. If not what's the point to set an affinity mask without a thread on it. Though the user is able to overwrite this by providing the -t options along side the -a option. > I'm not sure what you mean by this makes cyclictest behave the way the > rest of rt-tests does, the rest of rt-tests should match what cyclitest > does. As I wrote, make cyclictest behave the same way as the to other tools when -a/-t is used. > That said, I did some quick runs of signaltest and -a seems broken, > sigh. Is it broken without my patches or with my patches? What is broken?
On Tue, 26 Jan 2021, Daniel Wagner wrote: > On Tue, Jan 26, 2021 at 12:55:33AM -0500, John Kacur wrote: > > Well, --smp historically combined -a -t and threads at the same priority > > You could argue that it is reasonable for -a to automatically imply -t > > but I have had debates with people about this and we settled on -a just > > specifies the affinity, and the default number of threads is one unless > > you use -t. > > My thinking is if you set an affinity mask, you want also to assign a > thread to the CPUs. If not what's the point to set an affinity mask > without a thread on it. Though the user is able to overwrite this > by providing the -t options along side the -a option. Not necessarily, but you want to limit where the threads, no matter how few or how many run. As I said, it's not an unreasonable default, but it is not the current default of 1 unless you specify -t I suggest we stick with the current default in this clean-up effort. We can discuss if there are some better defaults after this effort is complete. > > > I'm not sure what you mean by this makes cyclictest behave the way the > > rest of rt-tests does, the rest of rt-tests should match what cyclitest > > does. > > As I wrote, make cyclictest behave the same way as the to other tools > when -a/-t is used. > > > That said, I did some quick runs of signaltest and -a seems broken, > > sigh. > > Is it broken without my patches or with my patches? What is broken? > > I haven't done a git bisect or something of that sort to determine if your patches broke it, but here's what I see. (running on non-rt laptop now, so ignore numbers), but it looks like at least the -t option requiring an argument existed before your changes, but your changes to affinity probably could use some more testing. The following is okay ./signaltest -a1-4 2.00 1.52 1.14 2/1079 6541 T: 0 ( 6537) P: 0 C: 14896 Min: 3 Act: 4 Avg: 26 Max: 366 The following is broken, the default without an argument should be 2 ./signaltest -t ./signaltest: option requires an argument -- 't' In the following note that threads other than thread 0 didn't appear on the screen until after the ctrl-c ./signaltest -t9 -a1-4 1.11 1.14 1.14 2/1075 6864 T: 0 ( 6856) P: 0 C: 1649 Min: 38 Act: 511 Avg: 457 Max: 4911 ^CT: 0 ( 6856) P: 0 C: 1660 Min: 38 Act: 554 Avg: 457 Max: 4911 T: 1 ( 6857) P: 0 C: 1660 Min: 38 Act: 490 Avg: 1090 Max: 15024 T: 2 ( 6858) P: 0 C: 1660 Min: 38 Act: 449 Avg: 1091 Max: 15053 T: 3 ( 6859) P: 0 C: 1660 Min: 38 Act: 412 Avg: 1091 Max: 15075 T: 4 ( 6860) P: 0 C: 1660 Min: 38 Act: 378 Avg: 1090 Max: 15081 T: 5 ( 6861) P: 0 C: 1659 Min: 38 Act: 517 Avg: 1091 Max: 15072 T: 6 ( 6862) P: 0 C: 1659 Min: 38 Act: 485 Avg: 1091 Max: 15067 T: 7 ( 6863) P: 0 C: 1659 Min: 38 Act: 522 Avg: 1091 Max: 12832 T: 8 ( 6864) P: 0 C: 1659 Min: 38 Act: 549 Avg: 1091 Max: 12839 etc, it's broken in many ways.
On Tue, Jan 26, 2021 at 11:32:25AM -0500, John Kacur wrote: > Not necessarily, but you want to limit where the threads, no matter > how few or how many run. As I said, it's not an unreasonable default, > but it is not the current default of 1 unless you specify -t > I suggest we stick with the current default in this clean-up effort. > We can discuss if there are some better defaults after this effort is > complete. Understood. BTW, you can do '-a 1 -t' which will spawn N threads on CPU 1. So it's still possible to get the same setup. > I haven't done a git bisect or something of that sort to determine if your > patches broke it, but here's what I see. (running on non-rt laptop now, so > ignore numbers), but it looks like at least the -t option requiring an > argument existed before your changes, but your changes to affinity > probably could use some more testing. > > The following is okay > > ./signaltest -a1-4 > 2.00 1.52 1.14 2/1079 6541 > > T: 0 ( 6537) P: 0 C: 14896 Min: 3 Act: 4 Avg: 26 Max: 366 > > > The following is broken, the default without an argument should be 2 > ./signaltest -t > ./signaltest: option requires an argument -- 't' Right, this shouldn't be too hard to fix. I'll send an update. > In the following note that threads other than thread 0 didn't appear > on the screen until after the ctrl-c > > ./signaltest -t9 -a1-4 > 1.11 1.14 1.14 2/1075 6864 > > T: 0 ( 6856) P: 0 C: 1649 Min: 38 Act: 511 Avg: 457 Max: 4911 > ^CT: 0 ( 6856) P: 0 C: 1660 Min: 38 Act: 554 Avg: 457 Max: 4911 > T: 1 ( 6857) P: 0 C: 1660 Min: 38 Act: 490 Avg: 1090 Max: 15024 > T: 2 ( 6858) P: 0 C: 1660 Min: 38 Act: 449 Avg: 1091 Max: 15053 > T: 3 ( 6859) P: 0 C: 1660 Min: 38 Act: 412 Avg: 1091 Max: 15075 > T: 4 ( 6860) P: 0 C: 1660 Min: 38 Act: 378 Avg: 1090 Max: 15081 > T: 5 ( 6861) P: 0 C: 1659 Min: 38 Act: 517 Avg: 1091 Max: 15072 > T: 6 ( 6862) P: 0 C: 1659 Min: 38 Act: 485 Avg: 1091 Max: 15067 > T: 7 ( 6863) P: 0 C: 1659 Min: 38 Act: 522 Avg: 1091 Max: 12832 > T: 8 ( 6864) P: 0 C: 1659 Min: 38 Act: 549 Avg: 1091 Max: 12839 > > > etc, it's broken in many ways. I don't think I broke this. This is since I can remember the output behavior. You only see the live update from T0, never the from the helper threads.
diff --git a/src/cyclictest/cyclictest.c b/src/cyclictest/cyclictest.c index 0e6519125c2f..b9b0eb3575e3 100644 --- a/src/cyclictest/cyclictest.c +++ b/src/cyclictest/cyclictest.c @@ -1024,6 +1024,8 @@ static void process_options(int argc, char *argv[], int max_cpus) argv[optind][0] == '0' || argv[optind][0] == '!')) { parse_cpumask(argv[optind], max_cpus, &affinity_mask); + } else { + num_threads = -1; } if (verbose)
Let --affinity without an argument behave in the same way as --smp. Run a thread on each CPU. This makes cyclictest behave as the rest of the rt-tests when --affinity is used. Reviewed-by: Daniel Wagner <dwagner@suse.de> --- src/cyclictest/cyclictest.c | 2 ++ 1 file changed, 2 insertions(+)