[go-nuts] Profiling output question

My particular piece of code runs about 3.5x faster (on my quad-core
machine) than in serial, so the concurrency seems to be pretty effective.
However, when I look at the CPU output from go tool pprof, the top couple
of items returned runtime waiting.

I don't understand how that can be. How can I have almost linear speedup
while half the time is spent waiting?


Parallel output
(pprof) top10
87.37s of 92.98s total (93.97%)
Dropped 215 nodes (cum <= 0.46s)
Showing top 10 nodes out of 73 (cum >= 0.91s)
       flat flat% sum% cum cum%
     54.94s 59.09% 59.09% 54.94s 59.09% runtime.mach_semaphore_wait
     18.35s 19.74% 78.82% 18.35s 19.74% runtime.mach_semaphore_timedwait
     10.92s 11.74% 90.57% 10.92s 11.74% runtime.usleep
      1.84s 1.98% 92.55% 1.84s 1.98% runtime.mach_semaphore_signal


Serial Output
(pprof) top5
18.96s of 79.70s total (23.79%)
Dropped 153 nodes (cum <= 0.40s)
Showing top 5 nodes out of 124 (cum >= 7.54s)
       flat flat% sum% cum cum%
      5.92s 7.43% 7.43% 11.94s 14.98% runtime.mallocgc
      3.84s 4.82% 12.25% 11.38s 14.28% runtime.makeslice
      3.46s 4.34% 16.59% 10.77s 13.51%
github.com/gonum/lapack/native.Implementation.Dlatrs
      3.32s 4.17% 20.75% 3.32s 4.17%
github.com/gonum/blas/native.Implementation.Dtrsv
      2.42s 3.04% 23.79% 7.54s 9.46% runtime.newarray