Re: [HACKERS] Profile of current backend

Can we go anywhere with this?

>
> Bruce Momjian wrote:
> >
> > Interesting.  Nothing is jumping out at me.  Looks like we could try to
> > clean up heapgettup() to see if there is anything in there that can be
> > speeded up.
> >
> > None of the calls looks like it should be inlined.  Do you see any that
> > look good for inlining?
>
> ExecScan() seems to be the only func which calls SeqNext(), which in
> turn accounts for 60% of the calls to heap_getnext(), which does 80% of
> the calls to heapgettup().
>
> - Put SeqNext() into ExecScan() to lower function call overhead? [minimal optim.]
>
> - In heapgettup(), 50% is the func itself and 50% is called funcs.
>   Top four CPU consumers:
>     0.04    0.14    9924/9924        RelationGetBufferWithBuffer [148]
>     0.03    0.15    5642/5702        ReleaseAndReadBuffer [145]
>     0.10    0.00   26276/42896       nocachegetattr [158]
>     0.01    0.08    7111/9607        HeapTupleSatisfiesVisibility [185]
>
>   RelationGetBufferWithBuffer() seems to be called from here only. If so, inline.
>
> - Looking at RelationGetBufferWithBuffer():
>     0.00    0.10    4603/32354       ReadBuffer [55]
>   ReadBuffer() is the biggest cpu consumer called by RelationGetBufferWithBuffer(). (55%)
>
>   -> *** 97% of ReadBuffer() CPU time is in calling ReadBufferWithBufferLock()
>
>   -> 85% of ReadBufferWithBufferLock() CPU time is in calling BufferAlloc().
>   -> ReadBufferWithBufferLock() is the only func calling BufferAlloc().
>   -> Conclusion: INLINE BufferAlloc().
>
> - Looking at BufferAlloc():
>     0.04    0.25   37974/37974       BufTableLookup [114]
>     0.10    0.00   32340/151781      SpinAcquire [81]
>     0.10    0.00   37470/40585       PinBuffer [209]
>     0.08    0.00   38478/43799       RelationGetLRelId [234]
>     0.04    0.00   37974/151781      SpinRelease [175]
>
>   -> 40% of BufferAlloc() CPU time is in calling BufTableLookup().
>   -> BufferAlloc() is the only func calling BufTableLookup().
>   -> Conclusion: INLINE BufTableLookup().
>
> - Looking at BufTableLookup():
>   86% of CPU time is in calling hash_search(). The rest is own time.
>
> - Looking at hash_search():
>     0.13    0.41  179189/179189      call_hash [69]
>     0.00    0.00       6/6           bucket_alloc [1084]
>   -> Conclusion: INLINE call_hash() [and bucket_alloc()] into hash_search().
>
> - Looking at call_hash():
>     0.37    0.00  171345/171345      tag_hash [94]
>     0.04    0.00    7844/7844        string_hash [348]
>   -> Conclusion: INLINE tag_hash() [and string_hash()] into call_hash().
>   -> Perhaps disk_hash() could be used in some way? It is currently #ifdef'd away.
>   -> Could we use a lookup table instead of doing hash calculations? Would not that
>   ->  be much faster?
>
>
> It looks to me as if there are too many levels of function calls.
> Perhaps all functions which are called by only one other func should be inlined?
>
>
> Guesstimate:
>   This would speed up heapgettup() by 10% ???
>   Other functions would speed up too.
>
>
> /* m */
>


--
Bruce Momjian                          |  830 Blythe Avenue
maillist@candle.pha.pa.us              |  Drexel Hill, Pennsylvania 19026
  +  If your life is a hard drive,     |  (610) 353-9879(w)
  +  Christ can be your backup.        |  (610) 853-3000(h)