Re: CLUSTER, reform_and_rewrite_tuple(), and parallelism

On Wed, Aug 17, 2016 at 4:28 PM, Andres Freund <andres@anarazel.de> wrote:
> Could you also provide a strace -ttt -T -c and a cpu cycles flamegraph?

Here is the output from that strace invocation, plus a -p (to attach
to the relevant backend):

strace: -t has no effect with -c
strace: -T has no effect with -c
strace: Process 27986 attached
strace: Process 27986 detached
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 55.75    0.629331       17981        35        16 unlink
 17.49    0.197422           0   2505449           write
 11.69    0.132000       11000        12           fsync
  8.13    0.091799           0   2078837           read
  5.32    0.060000       12000         5           ftruncate
  0.98    0.011011          24       460           brk
  0.64    0.007218        1805         4           munmap
  0.00    0.000050           0      6382           lseek
  0.00    0.000000           0        58         5 open
  0.00    0.000000           0        58           close
  0.00    0.000000           0        14           stat
  0.00    0.000000           0         4           mmap
  0.00    0.000000           0         2           rt_sigprocmask
  0.00    0.000000           0        12         6 rt_sigreturn
  0.00    0.000000           0         1           select
  0.00    0.000000           0        16           sendto
  0.00    0.000000           0         2         1 recvfrom
  0.00    0.000000           0        16           kill
  0.00    0.000000           0        19           semop
  0.00    0.000000           0        63           getrusage
  0.00    0.000000           0         5           epoll_create
  0.00    0.000000           0         9         4 epoll_wait
  0.00    0.000000           0        10           epoll_ctl
------ ----------- ----------- --------- --------- ----------------
100.00    1.128831               4591473        32 total

This doesn't seem that interesting, but not sure what you're looking for.

I also attach cycles flamegraph.

trace_sort indicated that the tuplesort CLUSTER takes just under 3
minutes (this includes writing out the new heap, of course).

--
Peter Geoghegan