tests/vm/fault_throughput.lua

   1 #!/usr/local/bin/recon
   2 require 'strict'
   3
   4 local benchrun = require 'benchrun'
   5 local perfdata = require 'perfdata'
   6 local csv = require 'csv'
   7 local sysctl = require 'sysctl'
   8 local os = require 'os'
   9
  10 local kDefaultDuration = 30
  11
  12 local benchmark = benchrun.new {
  13     name = 'xnu.zero_fill_fault_throughput',
  14     version = 1,
  15     arg = arg,
  16     modify_argparser = function(parser)
  17         parser:option{
  18             name = '--cpu-workers',
  19             description = 'Number of cpu workers'
  20         }
  21         parser:flag{
  22           name = '--through-max-workers',
  23           description = 'Run benchmark for [1..n] cpu workers'
  24         }
  25         parser:flag{
  26           name = '--through-max-workers-fast',
  27           description = 'Run benchmark for [1..2] and each power of four value in [4..n] cpu workers'
  28         }
  29         parser:option{
  30           name = '--path',
  31           description = 'Path to fault throughput binary'
  32         }
  33         parser:option{
  34           name = '--duration',
  35           description = 'How long, in seconds, to run each iteration',
  36           default = kDefaultDuration
  37         }
  38         parser:option{
  39             name = '--variant',
  40             description = 'Which benchmark variant to run (sparate-objects or share-objects)',
  41             default = 'separate-objects'
  42         }
  43     end
  44 }
  45
  46 assert(benchmark.opt.path, "No path supplied for fault throughput binary")
  47 assert(benchmark.opt.variant == "separate-objects" or
  48     benchmark.opt.variant == "share-objects", "Unsupported benchmark variant")
  49
  50 local ncpus, err = sysctl('hw.logicalcpu_max')
  51 assert(ncpus > 0, 'invalid number of logical cpus')
  52 local cpu_workers = tonumber(benchmark.opt.cpu_workers) or ncpus
  53
  54 local unit = perfdata.unit.custom('pages/sec')
  55 local tests = {}
  56
  57 function QueueTest(num_cores)
  58     table.insert(tests, {
  59         path = benchmark.opt.path,
  60         num_cores = num_cores,
  61     })
  62 end
  63
  64 if benchmark.opt.through_max_workers then
  65     for i = 1, cpu_workers do
  66         QueueTest(i)
  67     end
  68 elseif benchmark.opt.through_max_workers_fast then
  69     local i = 1
  70     while i <= cpu_workers do
  71         QueueTest(i)
  72         -- Always do a run with two threads to see what the first part of
  73         -- the scaling curve looks like
  74         -- (and to measure perf on dual core systems).
  75         if i == 1 and cpu_workers >= 2 then
  76             QueueTest(i + 1)
  77         end
  78         i = i * 4
  79     end
  80 else
  81     QueueTest(cpu_workers)
  82 end
  83
  84 for _, test in ipairs(tests) do
  85     local args = {test.path, "-v", benchmark.opt.variant, benchmark.opt.duration, test.num_cores,
  86                      echo = true}
  87     for out in benchmark:run(args) do
  88         local result = out:match("-----Results-----\n(.*)")
  89         benchmark:assert(result, "Unable to find result data in output")
  90         local data = csv.openstring(result, {header = true})
  91         for field in data:lines() do
  92             for k, v in pairs(field) do
  93                 benchmark.writer:add_value(k, unit, tonumber(v), {
  94                   [perfdata.larger_better] = true,
  95                   threads = test.num_cores,
  96                   variant = benchmark.opt.variant
  97                 })
  98             end
  99         end
 100     end
 101 end
 102
 103 benchmark:finish()