Marker API (CPU)

Close the connection to the LIKWID Marker API and write out measurement data to file. This file will be evaluated by likwid-perfctr.

getregion(regiontag::AbstractString, [num_events]) -> nevents, events, time, count

Get the intermediate results of the region identified by regiontag. On success, it returns * nevents: the number of events in the current group, * events: a list with all the aggregated event results, * time: the measurement time for the region and * count: the number of calls.

Initialize the Marker API, assuming that julia is running under likwid-perfctr. Must be called previous to all other functions.

init_dynamic(group_or_groups; kwargs...)

Initialize the full Marker API from within the current Julia session (i.e. no likwird-perfctr necessary). A performance group, e.g. "FLOPS_DP", must be provided as the first argument.

Initialize the Marker API only on the main thread (assuming that julia is running under likwid-perfctr). LIKWID.Marker.threadinit() must be called manually.

Checks whether the Marker API is active (by checking if the LIKWID_MODE environment variable has been set).

marker(f, regiontag::AbstractString)

Adds a LIKWID marker region around the execution of the given function f using Marker.startregion, Marker.stopregion under the hood. Note that LIKWID.Marker.init() and LIKWID.Marker.close() must be called before and after, respectively.

Examples

julia> using LIKWID

julia> Marker.init()

julia> marker("sleeping...") do
           sleep(1)
       end
true

julia> marker(()->rand(100), "create rand vec")
true

julia> Marker.close()

Switch to the next event set in a round-robin fashion. If you have set only one event set on the command line, this function performs no operation.

perfmon_marker(f, group_or_groups[; kwargs...])

Monitor performance groups in marked areas (see @marker) while executing the given function f on one or multiple Julia threads.

This is an experimental feature!

Note that

• Marker.init_dynamic, Marker.init, Marker.close, and PerfMon.finalize are called automatically
• the measurement of multiple performance groups is sequential and requires multiple executions of f!

Keyword arguments:

• cpuids (default: currently used CPU threads): specify the CPU threads (~ cores) to be monitored
• autopin (default: true): automatically pin Julia threads to the CPU threads (~ cores) they are currently running on (to avoid migration and wrong results).
• keep (default: false): keep the temporarily created marker file

Example

julia> using LIKWID

julia> perfmon_marker("FLOPS_DP") do
           # only the marked regions are monitored!
           NUM_FLOPS = 100_000_000
           a = 1.8
           b = 3.2
           c = 1.3
           @marker "calc_flops" for _ in 1:NUM_FLOPS
                c = a * b + c
            end
           z = a*b+c
           @marker "exponential" exp(z)
           sin(c)
       end

Region: calc_flops, Group: FLOPS_DP
┌───────────────────────────┬───────────┐
│                     Event │  Thread 1 │
├───────────────────────────┼───────────┤
│          ACTUAL_CPU_CLOCK │ 3.00577e8 │
│             MAX_CPU_CLOCK │ 2.08917e8 │
│      RETIRED_INSTRUCTIONS │ 3.00005e8 │
│       CPU_CLOCKS_UNHALTED │ 3.00067e8 │
│ RETIRED_SSE_AVX_FLOPS_ALL │     1.0e8 │
│                     MERGE │       0.0 │
└───────────────────────────┴───────────┘
┌──────────────────────┬───────────┐
│               Metric │  Thread 1 │
├──────────────────────┼───────────┤
│  Runtime (RDTSC) [s] │ 0.0852431 │
│ Runtime unhalted [s] │  0.122687 │
│          Clock [MHz] │   3524.84 │
│                  CPI │   1.00021 │
│         DP [MFLOP/s] │   1173.12 │
└──────────────────────┴───────────┘

Region: exponential, Group: FLOPS_DP
┌───────────────────────────┬──────────┐
│                     Event │ Thread 1 │
├───────────────────────────┼──────────┤
│          ACTUAL_CPU_CLOCK │  85696.0 │
│             MAX_CPU_CLOCK │  59192.0 │
│      RETIRED_INSTRUCTIONS │   5072.0 │
│       CPU_CLOCKS_UNHALTED │   6013.0 │
│ RETIRED_SSE_AVX_FLOPS_ALL │     27.0 │
│                     MERGE │      0.0 │
└───────────────────────────┴──────────┘
┌──────────────────────┬────────────┐
│               Metric │   Thread 1 │
├──────────────────────┼────────────┤
│  Runtime (RDTSC) [s] │ 2.60005e-7 │
│ Runtime unhalted [s] │ 3.49786e-5 │
│          Clock [MHz] │    3546.95 │
│                  CPI │    1.18553 │
│         DP [MFLOP/s] │    103.844 │
└──────────────────────┴────────────┘

Register a region with name regiontag to the Marker API. On success, true is returned.

This is an optional function to reduce the overhead of region registration at Marker.startregion. If you don't call registerregion, the registration is done at startregion.

Reset the values stored using the region name regiontag. On success, true is returned.

Start measurements under the name regiontag. On success, true is returned.

Stop measurements under the name regiontag. On success, true is returned.

Add the current thread to the Marker API.

Convenience macro for flanking code with Marker.startregion and Marker.stopregion.

Examples

julia> using LIKWID

julia> Marker.init()

julia> @marker "sleeping..." sleep(1)
true

julia> @marker "create rand vec" rand(100)
true

julia> Marker.close()

Convenience macro for flanking code with Marker.startregion and Marker.stopregion on all threads separately.

Examples

julia> using LIKWID

julia> Marker.init()

julia> @parallelmarker begin
           Threads.@thread :static for i in 1:Threads.nthreads()
               # thread-local computation
           end
       end

julia> Marker.close()

@perfmon_marker group_or_groups codeblock

This is an experimental feature!

See also: perfmon_marker

Example

julia> using LIKWID

julia> @perfmon_marker "FLOPS_DP" begin
           @marker "exponential" exp(3.141)
       end

Region: exponential, Group: FLOPS_DP
┌───────────────────────────┬──────────┐
│                     Event │ Thread 1 │
├───────────────────────────┼──────────┤
│          ACTUAL_CPU_CLOCK │ 115146.0 │
│             MAX_CPU_CLOCK │  78547.0 │
│      RETIRED_INSTRUCTIONS │   4208.0 │
│       CPU_CLOCKS_UNHALTED │   7112.0 │
│ RETIRED_SSE_AVX_FLOPS_ALL │     10.0 │
│                     MERGE │      0.0 │
└───────────────────────────┴──────────┘
┌──────────────────────┬────────────┐
│               Metric │   Thread 1 │
├──────────────────────┼────────────┤
│  Runtime (RDTSC) [s] │ 3.02056e-8 │
│ Runtime unhalted [s] │ 4.70008e-5 │
│          Clock [MHz] │     3591.4 │
│                  CPI │    1.69011 │
│         DP [MFLOP/s] │    331.064 │
└──────────────────────┴────────────┘