Re: [Tarantool-patches] [PATCH v1 luajit 04/41] perf: adjust array3d in LuaJIT-benches

[Sergey Kaplun via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Hi, Sergey!
Thanks for the review!
Please consider my comments below.

On 13.11.25, Sergey Bronnikov wrote:
> Hi, Sergey!
> 
> thanks for the patch!
> 
> On 10/24/25 13:50, Sergey Kaplun wrote:

<snipped>

Also, added the comment with the benchmark description as we discussed
offline.

===================================================================
diff --git a/perf/LuaJIT-benches/array3d.lua b/perf/LuaJIT-benches/array3d.lua
index 1062fff0..4d88432d 100644
--- a/perf/LuaJIT-benches/array3d.lua
+++ b/perf/LuaJIT-benches/array3d.lua
@@ -1,3 +1,7 @@
+-- The benchmark to measure simple operations on array-like data
+-- structures.
+-- Here, a 3D array is represented as a contiguous 1D array.
+
 local bench = require("bench").new(arg)
 
 local function array_set(self, x, y, z, p)
===================================================================

> > -local arr = array_new(dim, dim, dim, packed)
> >   
> > -for x,y,z inarr:points() do
> > -arr:set(x, y, z, x*x)
> > -end
> > -assert(arr.image[dim^3-1] == (dim-1)^2)
> > +bench:add({
> > +  name = "array3d",
> > +  checker = function(arr)
> > +    assert(arr.image[dim^3-1] == (dim-1)^2)
> > +    return true
> > +  end,
> > +  payload = function()
> > +    local arr = array_new(dim, dim, dim, packed)
> > +    for x,y,z inarr:points() do
> please add whitespaces after commas

Added:

===================================================================
diff --git a/perf/LuaJIT-benches/array3d.lua b/perf/LuaJIT-benches/array3d.lua
index 75ab5b01..1062fff0 100644
--- a/perf/LuaJIT-benches/array3d.lua
+++ b/perf/LuaJIT-benches/array3d.lua
@@ -60,7 +60,7 @@ bench:add({
   end,
   payload = function()
     local arr = array_new(dim, dim, dim, packed)
-    for x,y,z in arr:points() do
+    for x, y, z in arr:points() do
       arr:set(x, y, z, x*x)
     end
     return arr
===================================================================

But I'm not sure that we need to refactor each part of the code this
way. Anyway, I'm OK with reformatting this since we already touched this
chunk.

> > +arr:set(x, y, z, x*x)
> > +    end
> > +    return arr
> > +  end,
> > +  items = dim * dim * dim,
> > +  -- Limit the number of iterations to avoid OOM errors for
> > +  -- non-GC64 builds.
> > +  iterations = 5,

Replaced to 4 iterations, since sometimes it strikes the OOM on the
runner.

===================================================================
diff --git a/perf/LuaJIT-benches/array3d.lua b/perf/LuaJIT-benches/array3d.lua
index 4d88432d..80562706 100644
--- a/perf/LuaJIT-benches/array3d.lua
+++ b/perf/LuaJIT-benches/array3d.lua
@@ -72,7 +72,7 @@ bench:add({
   items = dim * dim * dim,
   -- Limit the number of iterations to avoid OOM errors for
   -- non-GC64 builds.
-  iterations = 5,
+  iterations = 4,
 })
 
 bench:run_and_report()
===================================================================

> > +})
> >   
> > +bench:run_and_report()
> 
> Looks like benchmark min time does not work as expected:
> 
> [1] ~/sources/MRG/tarantool/third_party/luajit $ time ./build/src/luajit 
> perf/LuaJIT-benches/array3d.lua --benchmark_min_time=10
> -------------------------------------------------------------------------------------------------------------
> Benchmark                                     Time  CPU    Iterations 
> UserCounters...
> -------------------------------------------------------------------------------------------------------------
> array3d                                     2.10 s          2.13 s      
>         5 items_per_second=64.370M/s
> 
> real    0m2.333s
> user    0m1.869s
> sys     0m0.461s
> [1] ~/sources/MRG/tarantool/third_party/luajit $
> 
> --benchmark_min_time set to 10 sec, but benchmark.lua reports "2.10 s" 
> and time reported by `time` utility
> 
> is less than 10 sec.

It works exactly as expected, since this option is superseded by the
`iterations` value in the becnhmark configuration, see the corresponding
comment in the code for the rationale in this particular benchmark.
Also, it is mentioned in the commit message.

-- 
Best regards,
Sergey Kaplun