Messages in this thread Patch in this message | | | Date | Tue, 19 May 2015 16:27:02 -0600 | From | David Ahern <> | Subject | Re: [PATCH 38/40] perf session: Handle index files generally |
| |
On 5/17/15 6:30 PM, Namhyung Kim wrote: > The current code assumes that the number of index item and cpu are > matched so it creates that number of threads. But it's not the case > of non-system-wide session or data came from different machine. > > Just creates threads at most number of online cpus and process data.
-----8<-----
> @@ -1717,6 +1742,7 @@ int perf_session__process_events_mt(struct perf_session *session, void *arg) > int err, i, k; > int nr_index = session->header.nr_index; > u64 size = perf_data_file__size(file); > + int nr_thread = sysconf(_SC_NPROCESSORS_ONLN);
It's not clear to me how this multi-threaded perf is going to work on large systems especially this patch if a system has holes in the active cpus. e.g,
# lscpu Architecture: sparc64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Big Endian CPU(s): 704 On-line CPU(s) list: 32-63,128-223,256-351,384-479,576-831,864-927,960-1023 Thread(s) per core: 12 Core(s) per socket: 18 Socket(s): 3 NUMA node(s): 4 NUMA node0 CPU(s): 32-63,128-223 NUMA node1 CPU(s): 256-351,384-479 NUMA node2 CPU(s): 576-767 NUMA node3 CPU(s): 768-831,864-927,960-1023
So you are going to spawn 704 threads? Each thread handles a per-cpu buffer?
yes, I still need to find time to take if for a test drive; maybe by the end of the week.
David
| |