[LLVMdev] On LLD performance

[Rafael Espíndola]

> The biggest difference that shows up is that lld has 1,152 context
> switches, but the cpu utilization is still < 1. Maybe there is just a
> threading bug somewhere?

It was a bug, but in my script: I was pining the process to a single
core, a leftover from benchmarking single threaded code. sorry about
that.

With that fixed (new script attached) what I got when linking clang
with a today's build of clang, lld and gold (so not directly
comparable to the old numbers) was:


lld:

       2544.422504      task-clock (msec)         #    1.837 CPUs
utilized            ( +-  0.13% )
            69,815      context-switches          #    0.027 M/sec
               ( +-  0.10% )
               515      cpu-migrations            #    0.203 K/sec
               ( +-  2.38% )
           188,589      page-faults               #    0.074 M/sec
               ( +-  0.00% )
     7,518,975,892      cycles                    #    2.955 GHz
               ( +-  0.12% )
     4,848,133,388      stalled-cycles-frontend   #   64.48% frontend
cycles idle     ( +-  0.18% )
   <not supported>      stalled-cycles-backend
     6,383,065,952      instructions              #    0.85  insns per
cycle
                                                  #    0.76  stalled
cycles per insn  ( +-  0.06% )
     1,331,202,027      branches                  #  523.184 M/sec
               ( +-  0.07% )
        30,053,442      branch-misses             #    2.26% of all
branches          ( +-  0.04% )

       1.385020712 seconds time elapsed
          ( +-  0.15% )

gold:

        918.859273      task-clock (msec)         #    0.999 CPUs
utilized            ( +-  0.01% )
                 0      context-switches          #    0.000 K/sec
               ( +- 41.02% )
                 0      cpu-migrations            #    0.000 K/sec
            76,393      page-faults               #    0.083 M/sec
     2,758,439,113      cycles                    #    3.002 GHz
               ( +-  0.01% )
     1,357,937,367      stalled-cycles-frontend   #   49.23% frontend
cycles idle     ( +-  0.02% )
   <not supported>      stalled-cycles-backend
     3,881,565,068      instructions              #    1.41  insns per
cycle
                                                  #    0.35  stalled
cycles per insn  ( +-  0.00% )
       784,078,474      branches                  #  853.317 M/sec
               ( +-  0.00% )
        13,984,077      branch-misses             #    1.78% of all
branches          ( +-  0.01% )

       0.919717669 seconds time elapsed
          ( +-  0.01% )

gold --threads

       1300.210314      task-clock (msec)         #    1.523 CPUs
utilized            ( +-  0.15% )
            26,913      context-switches          #    0.021 M/sec
               ( +-  0.34% )
             7,910      cpu-migrations            #    0.006 M/sec
               ( +-  0.63% )
            83,507      page-faults               #    0.064 M/sec
               ( +-  0.02% )
     3,842,183,459      cycles                    #    2.955 GHz
               ( +-  0.14% )
     2,273,634,375      stalled-cycles-frontend   #   59.18% frontend
cycles idle     ( +-  0.20% )
   <not supported>      stalled-cycles-backend
     4,058,907,107      instructions              #    1.06  insns per
cycle
                                                  #    0.56  stalled
cycles per insn  ( +-  0.07% )
       838,791,061      branches                  #  645.120 M/sec
               ( +-  0.06% )
        14,220,563      branch-misses             #    1.70% of all
branches          ( +-  0.02% )

       0.853835099 seconds time elapsed
          ( +-  0.25% )

So much better!

The machine has 12 cores, each with 2 hardware threads.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: run.sh
Type: application/x-sh
Size: 5352 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150319/5f2c7cae/attachment.sh>