Messages in this thread | | | Date | Tue, 31 Oct 2017 12:26:57 -0300 | From | Arnaldo Carvalho de Melo <> | Subject | Re: [PATCH v3 2/2] perf/bench/numa: Handle discontiguous/sparse numa nodes |
| |
Em Tue, Oct 31, 2017 at 08:46:58PM +0530, Naveen N. Rao escreveu: > On 2017/08/21 10:17AM, sathnaga@linux.vnet.ibm.com wrote: > > From: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com> > > > > Certain systems are designed to have sparse/discontiguous nodes. > > On such systems, perf bench numa hangs, shows wrong number of nodes > > and shows values for non-existent nodes. Handle this by only > > taking nodes that are exposed by kernel to userspace. > > > > Cc: Arnaldo Carvalho de Melo <acme@kernel.org> > > Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> > > Signed-off-by: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com> > > Signed-off-by: Balamuruhan S <bala24@linux.vnet.ibm.com> > > --- > > tools/perf/bench/numa.c | 17 ++++++++++------- > > 1 file changed, 10 insertions(+), 7 deletions(-) > > > > diff --git a/tools/perf/bench/numa.c b/tools/perf/bench/numa.c > > index 2483174..d4cccc4 100644 > > --- a/tools/perf/bench/numa.c > > +++ b/tools/perf/bench/numa.c > > @@ -287,12 +287,12 @@ static cpu_set_t bind_to_cpu(int target_cpu) > > > > static cpu_set_t bind_to_node(int target_node) > > { > > - int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes; > > + int cpus_per_node = g->p.nr_cpus/nr_numa_nodes(); > > cpu_set_t orig_mask, mask; > > int cpu; > > int ret; > > > > - BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus); > > + BUG_ON(cpus_per_node*nr_numa_nodes() != g->p.nr_cpus); > > BUG_ON(!cpus_per_node); > > > > ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask); > > @@ -692,7 +692,7 @@ static int parse_setup_node_list(void) > > int i; > > > > for (i = 0; i < mul; i++) { > > - if (t >= g->p.nr_tasks) { > > + if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) { > > printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node); > > goto out; > > } > > @@ -973,6 +973,7 @@ static void calc_convergence(double runtime_ns_max, double *convergence) > > int node; > > int cpu; > > int t; > > + int processes; > > > > if (!g->p.show_convergence && !g->p.measure_convergence) > > return; > > @@ -1007,13 +1008,14 @@ static void calc_convergence(double runtime_ns_max, double *convergence) > > sum = 0; > > > > for (node = 0; node < g->p.nr_nodes; node++) { > > + if (!is_node_present(node)) > > + continue; > > nr = nodes[node]; > > nr_min = min(nr, nr_min); > > nr_max = max(nr, nr_max); > > sum += nr; > > } > > BUG_ON(nr_min > nr_max); > > - > > Looks like an un-necessary change there.
Right, and I would leave the 'int processes' declaration where it is, as it is not used outside that loop.
The move of that declaration to the top of the calc_convergence() function made me spend some cycles trying to figure out why that was done, only to realize that it was an unnecessary change :-\
> - Naveen > > > BUG_ON(sum > g->p.nr_tasks); > > > > if (0 && (sum < g->p.nr_tasks)) > > @@ -1027,8 +1029,9 @@ static void calc_convergence(double runtime_ns_max, double *convergence) > > process_groups = 0; > > > > for (node = 0; node < g->p.nr_nodes; node++) { > > - int processes = count_node_processes(node); > > - > > + if (!is_node_present(node)) > > + continue; > > + processes = count_node_processes(node); > > nr = nodes[node]; > > tprintf(" %2d/%-2d", nr, processes); > > > > @@ -1334,7 +1337,7 @@ static void print_summary(void) > > > > printf("\n ###\n"); > > printf(" # %d %s will execute (on %d nodes, %d CPUs):\n", > > - g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus); > > + g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus); > > printf(" # %5dx %5ldMB global shared mem operations\n", > > g->p.nr_loops, g->p.bytes_global/1024/1024); > > printf(" # %5dx %5ldMB process shared mem operations\n",
| |