Using Backends

In this section you can find useful information of the different backends currently implemented.

To use them just specify the NESMIK_BACKEND environment variable to the case-sensitive backend name. E.g. NESMIK_BACKEND=Default

Note

TL;DR

NESMIK_BACKEND=

Requires

Short Description

Default

Prints every region start and stop to stdout.

Detection

Helps identify problems with the annotation.

Extrae::TypeStack

EXTRAE

Creates paraver events for each annotation.

TALP

DLB

Simple wrapper around TALP monitoring regions.

TALP::Tree

DLB

Call-stack sensitive wrapper to TALP monitoring regions.

CaPI

DLB

Uses hash values of callstack as region identifiers.

CaPIColl

DLB

Like CaPI but detects hash collisions.

NSYS

CUDA

Can control the NVIDIA nsys tracer

OTF-CPT

MPI

Uses MPI Pcontrol to control the OTF-CPT

SCOREP

MPI

Emits anonymous SCOREP regions

MERIC

MERIC

Wrapper around MERIC monitoring regions

Independent backends

These backends don’t require any dependency and are always provided, and are helpful to solve problems that you might encounter with the annotation.

Default

The Default backend implements both Properly Nested and Not Properly Nested versions. This backend will just print the region name to the stdout and can be used to check if neSmiK is working correctly.

Detection

This backend helps users to identify problems with their annotation. It checks the annotations if they are actually properly nested or not and tries to help with problematic regions that are not properly nested. It also checks if annoatations are called from multiple threads, as not all backends support that.

Options

  • NESMIK_DETECTION_SHOW_ALL_REGIONS Report summary for all regions. If disabled, only regions with odd behaviour will be reported. (default:off)”

Extrae Backends

Note

Requires neSmiK to be configured with -DWITH_EXTRAE=ON

To use the Extrae backends you need to set up an Extrae run as usual, and also set the NESMIK_BACKEND. Here is an example for the Extrae::TypeStack backend:

export NESMIK_BACKEND="Extrae::TypeStack"
export EXTRAE_CONFIG_FILE="./extrae.xml" # where you put you extrae configuration
mpirun -n <nprocs> env LD_PRELOAD=${EXTRAE_HOME}/lib/libmpitrace.so ./test

Extrae::TypeStack

The Extrae::TypeStack provides a way to get your annotations into paraver.

Options

  • NESMIK_EXTRAE_WRITE_CONFIG_FILE When setting this to a boolean variable (1,0) the backend will generate corresponding .cfg files to view the anntation in the directory the application was run from.

Post-process traces

We recommend to use pertalde to postprocess the generated trace, as the semantic values of the region string are their hash values.

Using pertalde unir -p nesmik <trace.prv> will post-process the trace to make all the values of neSmiK regions start at 1, and increment them for each semantic, which is highly beneficial as Paraver does not support arbitrary value ranges for histograms. Additionally, if you have multiple traces you can unify them all with the same command, which will ensure that all the traces share the values for the same semantics, for example pertalde unir -p nesmik trace1.prv trace2.prv.

DLB Backends

Note

Requires neSmiK to be configured with -DWITH_DLB=ON

To use the DLB backends you still need to configure your execution to run with DLB. This requires preloading the right DLB library and setting the correct DLB_ARGS.

For example, for the TALP backend:

export NESMIK_BACKEND="TALP"
export DLB_ARGS="--talp"
mpirun -n <nprocs> env LD_PRELOAD=${DLB_HOME}/lib/libdlb_mpi.so ./test

TALP

The TALP backend is a simple wrapper into the TALP module of the DLB Library.

TALP::Tree

The TALP::Tree backend is a call path sensitive wrapper of TALP monitoring regions. This backend will create a monitoring region for each unique hierrarchy of regions.

CaPI

The CaPI backend uses the hash value of the callstack as the region identifier. This is useful when you have many different regions with the same name but different call paths, as TALP will treat each unique call stack as a separate region. Using a special toolchain an annotated metacg callgraph can be resconstructed and used with other CAPI based tools.

CaPIColl

The CaPIColl backend is identical to the CaPI backend (a wrapper around TALP that uses hash values of the callstack as region identifiers), but with additional collision detection. It monitors for hash collisions where different call stacks produce the same hash value.

When collisions are detected, they are reported in the console output from MPI rank 0. This backend is useful for debugging and verifying whether hash collisions could be affecting the accuracy when using the CaPI backend.

Options

  • NESMIK_JSON_OUTPUT: Enable output to a JSON file. (default: on)

  • NESMIK_JSON_FILE: Set the name of the output JSON file. (default: capinesmik_collision_detection.json)

Nvidia Nsight Systems (nsys) Backends

Note

Requires neSmiK to be configured with -DENABLE_NSYS=ON

NSYS

The NSYS backend is a simple wrapper into the NVTX Library. All the neSmiK annotations are separated into a domain called neSmiK, to not interfere with other annotation libs.

OTF-CPT Backend

Note

Requires neSmiK to be configured with -DENABLE_OTFCPT=ON and -DWITH_MPI=ON

The OTF-CPT backend writes traces in OTF-CPT format, which can be analyzed with the CPS Tool.

To use:

export NESMIK_BACKEND="OTF-CPT"
mpirun -n <nprocs> ./test

SCOREP Backend

Note

Requires neSmiK to be configured with -DENABLE_SCOREP=ON and -DWITH_MPI=ON

The SCOREP backend integrates with the SCOREP profiling infrastructure. All neSmiK regions are recorded as SCOREP user regions.

To use:

export NESMIK_BACKEND="ScoreP"
mpirun -n <nprocs> ./test

MERIC Backend

Note

Requires neSmiK to be configured with -DENABLE_MERIC=ON and -DMERIC_DIR=<path/to/meric>

The MERIC backend integrates with the MERIC performance measurement infrastructure. MERIC can measure energy and performance of specific code regions.

To use:

export NESMIK_BACKEND="MERIC"
./test

  • NESMIK_MERIC_ENUMERATE: Comma-separated list of region names to enumerate. When set, MERIC will create unique measurements for each invocation of the specified regions by appending ___0, ___1, etc. to the region name. This is useful for measuring multiple iterations of the same region separately.

    Example:

    export NESMIK_MERIC_ENUMERATE="region1,region2"
./test