Re: + lib-vsprintfc-even-faster-decimal-conversion.patch added to -mm tree

[Rasmus Villemoes]

On Fri, Mar 13 2015, Alexey Dobriyan <adobriyan@xxxxxxxxx> wrote:

> On Thu, Mar 12, 2015 at 12:54 AM,  <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>> Subject: lib/vsprintf.c: even faster binary to decimal conversion
>
> I spent some time to microbenchmark changes in userspace (audience: fool!).
> Results are below.
>
> Legend is "number avg+-1sigma min-max". Every number is CPU cycles.
> Great care was taken to remove interrupt noise.
>
> Number of measurements is 100 millions per line.
> CPU is Intel Core 2 Duo E6550 in 64-bit mode.
>
> 3.19.1:
>
> 0                     98.015369 +- 0.512937   91-616
> 42                   116.000193 +- 3.523826  112-868
> 27182                137.009008 +- 3.515725  133-1043
> 65535                137.008262 +- 3.521761  133-840
> 4294967295           201.019966 +- 3.278608  196-1050
> 3141592653589793238  289.996882 +- 3.489376  287-1148
> 18446744073709551615 295.065274 +- 2.860187  287-1029
> -----------------------------------------------------
> 3.19.1+patch
> 0                     94.444063 +- 3.518922   84-630
> 42                   116.428533 +- 18.539093 105-1036
> 42                   116.316904 +- 18.234484 105-833
> 27182                136.172398 +- 3.737113  133-980
> 65535                136.014742 +- 3.537882  133-714
> 4294967295           172.009618 +- 3.507473  168-826
> 3141592653589793238  207.001114 +- 3.492724  196-1120
> 18446744073709551615 208.018154 +- 3.220185  203-1246
> -----------------------------------------------------

This seems to measure lfence+rdtsc overhead more than anything else. On
my presumably rather similar Core2 Duo T5870, I get an average of 22
cycles for the old code and 11 cycles for the new when converting 42 two
million times in a loop, and I'm not even trying to take interrupts into
account.

Since you seem to be dirtying 800 MB of memory, I'm guessing you get
quite a few page faults/TLB misses, which might explain the rather huge
max numbers.

> New code is somewhat faster for huge numbers.
> But top and ps don't show huge numbers normally --
> it is either PIDs (2^16) or moderately high numbers in a range of millions
> (see /proc/stat)

I said much the same thing in the commit log, and accordingly I've done
(micro)benchmarks with distributions biased to various degrees towards
smaller numbers, all of which showed 25+% improvement.

> * variance for new code is bigger

Seems to depend on how you measure...

> I even tried N=42 twice because I thought 18.5 variance is a glitch
> but it is not.

That does seem odd. But I think your numbers are caused by the huge
memory use. In any case, I modified my test program to record the
cycle count for each individual call (using lfence+rdtsc), but I used a
frequency table instead of a gigantic array, ignoring excessively large
cycle counts (I used > 1023). On the Core 2, I then get

90 90 90 80 80 80 80 80 80 80 80 80 90 80 80
Distribution              Function         cycles/conv  std.dev. (ignored) 
uniform([10, 2^64-1])     linux_put_dec          224.83     9.80 (156)
uniform([10, 2^64-1])     rv_put_dec             147.87     7.44 (86)
3 + neg_binom(0.05)       linux_put_dec          138.97    39.48 (87)
3 + neg_binom(0.05)       rv_put_dec             123.76    27.33 (77)
3 + neg_binom(0.10)       linux_put_dec          115.49    27.27 (84)
3 + neg_binom(0.10)       rv_put_dec             108.22    20.14 (71)
3 + neg_binom(0.15)       linux_put_dec          105.21    20.52 (59)
3 + neg_binom(0.15)       rv_put_dec             101.75    17.21 (54)
3 + neg_binom(0.20)       linux_put_dec          100.79    17.25 (65)
3 + neg_binom(0.20)       rv_put_dec              98.34    16.22 (64)
3 + neg_binom(0.50)       linux_put_dec           87.84     7.75 (44)
3 + neg_binom(0.50)       rv_put_dec              85.37     8.26 (45)

[first line is just deltas between a few lfence+rdtsc reads in quick
succession, to get a sense of the overhead]. For each distribution I'm
generating 2048 random numbers and then iterate over that 1000 times. So
almost none of the ~2M observations are being ignored. Here, the new
code is always faster (but computing a percentage from numbers including
the rdtsc overhead is meaningless), and in all but the last case (where
the numbers are almost exclusively 2-digit) the std. deviation is also
smaller. An an Intel Xeon, I get

48 32 32 32 32 32 32 32 32 32 32 32 32 32 32
Distribution              Function         cycles/conv  std.dev. (ignored) 
uniform([10, 2^64-1])     linux_put_dec          152.58     8.54 (26)
uniform([10, 2^64-1])     rv_put_dec              89.33     3.02 (16)
3 + neg_binom(0.05)       linux_put_dec           91.88    34.46 (17)
3 + neg_binom(0.05)       rv_put_dec              71.59    21.19 (10)
3 + neg_binom(0.10)       linux_put_dec           72.50    25.30 (12)
3 + neg_binom(0.10)       rv_put_dec              60.10    17.51 (11)
3 + neg_binom(0.15)       linux_put_dec           63.81    20.68 (8)
3 + neg_binom(0.15)       rv_put_dec              55.57    15.74 (6)
3 + neg_binom(0.20)       linux_put_dec           57.18    16.50 (7)
3 + neg_binom(0.20)       rv_put_dec              51.15    13.58 (12)
3 + neg_binom(0.50)       linux_put_dec           45.06     6.39 (4)
3 + neg_binom(0.50)       rv_put_dec              41.16     6.51 (6)

> New code uses lookup table which implies cache misses.  Current code
> is purely code.

Code can miss the cache also, and then it needs to be decoded again. The
new _code_ is slightly smaller, although the total .text+.rodata does
increase by ~150 bytes. Yes, overall the new code will probably
touch one or two extra cache lines compared to the old - so there's a
tradeoff between one-shot and bulk decimal conversions.

>> On a larger scale, perf shows that top, one of the big consumers of /proc
>> data, uses 0.5-1.0% fewer cpu cycles.
>
> perf(1) also shows variance next to average, what was it?

Not in the output I got - that just showed lines such as 

     2.35%  top      [kernel.kallsyms]   [k] num_to_str                 

But I don't have much perf-fu, so maybe I should have invoked it differently.

> First number printing improvement patch was measuring ~30% speedups:
> commit 4277eedd7908a0ca8b66fad46ee76b0ad96e6ef2
> vsprintf.c: optimizing, part 2: base 10 conversion speedup, v2
>
> Now it is 1%.

That's comparing apples and oranges. The ~30% were obtained from a top
which was modified to do nothing but read /proc/pid/stat, the 1% is out
of all the cycles top actually spends.

> I think any further improvements to number printing code should be rejected
> on philosophical grounds:

Hm, "perfect is the enemy of good" and all that.

> Kernel should ship numbers to ps(1) and top(1) in BINARY,
> so it would take exactly 1 MOV instruction which takes exactly 1 cycle
> to execute.
> Currently it is 1) kernel converts binary to text, 2) usespace
> converts text to binary,
> 3) userspace converts binary to text and shows the user. 4) people optimizing #1

I agree this is somewhat silly, but that's what we have, and it is
unlikely to change anytime soon. Adding a parallel binary interface
would be a maintenance nightmare.

Rasmus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/