This repository holds the anomaly suite presented in the paper "HPAS: An HPC Performance Anomaly Suite for Reproducing Performance Variations"[1]. It consists of a set of synthetic anomalies that reproduce common root causes of performance variations in supercomputers:

• CPU contention
• Cache evictions
• Memory bandwidth interference
• Memory intensive processes
• Memory leaks
• Network contention
• I/O metadata server contention
• I/O storage server contention

These anomalies use processes that run in user space; thus do not require any hardware or kernel modification. For installation details please see below.

An earlier subset of the anomalies were used as part of the papers Tuncer et al., "Online Diagnosis of Performance Variation in HPC Systems Using Machine Learning"[2] and Tuncer et al., "Diagnosing Performance Variations in HPC Applications Using Machine Learning"[3].

A brief intro is here, but check INSTALL for more details.

1. Install C/C++ compilers, GNU autotools
   1. Install shmem if you wish to use netoccupy anomaly. OpenMPI typically includes a shmem compiler.
2. If checking out from git, generate the configure script by running the command ./autogen.sh. For the latest release tarball, see the releases.
3. ./configure
   1. Configure has some options that can be explored, such as prefix directory, see ./configure --help
   2. The cache anomalies measure the cache size during this step, so it's important to run configure on the node that the anomaly is going to execute on.
   3. LD_LIBRARY_PATH and CFLAGS should indicate the location of shmem.h and relevant libraries. If using OpenMPI, use oshcc to compile (add CC=oshcc to configure arguments) and oshrun to run.
4. make
5. make install

The time units are specified in double precision floats, in seconds. Up to microsecond granularity can be used. Each anomaly has a --help option, which prints detailed information about command line arguments.

For detailed evaluation of anomalies, refer to our paper at ICPP'19, a copy is included in docs/ates_icpp19.pdf.

The netoccupy anomaly assumes that each node has a hostname of the form e.g., nid00020, such that each node name is nid followed by a unique integer. The relevant line within src/netoccupy.c should be changed to support different node naming conventions.

In the 2017 ISC paper [3], and the 2019 TPDS paper [2] an earlier version of the anomaly suite was used. The following table describes how each anomaly corresponds to the ones in the suite.

In [2], linkclog was using wrappers around MPI calls to emulate network contention, whereas netoccupy creates actual network contention by sending/receiving many messages, so they are not equivalent.

For SLURM systems, the anomalies can be injected into the application with the following example line (with necessary modifications) in the job submission script:

srun --nodelist=$ANOMALOUS_NODE -N 1 -n 1 $PREFIX/bin/hpas cpuoccupy [--options] &
anomaly_pid=$!

For anomalies that stress a shared resource between nodes (I/O and network), $ANOMALOUS_NODE should be different from application nodes, and it should be one of the application nodes if the anomaly is a CPU, cache or memory anomaly. The I/O anomalies can be executed with higher -N (nodes) and -n (ranks) values for more interference, and the network anomaly has to be executed with -N 2.

This anomaly suite is not an exhaustive list of anomaly generators for HPC systems. Therefore, we are open to additions of new anomalies. If you have an anomaly generator you would like to contribute, please open a pull request/issue at GitHub, and we can work together to include the anomalies.

For any bugs/problems, please open a pull request if you have a fix, issue otherwise.

If you use this anomaly suite for a publication, please cite [1].

Bibtex entry:

@inproceedings{ates:2019,
    author = {Emre Ates and Yijia Zhang and Burak Aksar and Jim Brandt and
              Vitus J. Leung and Manuel Egele and Ayse K. Coskun},
    title = {{HPAS}: An {HPC} Performance Anomaly Suite for Reproducing
             Performance Variations},
    booktitle = {48th International Conference on Parallel Processing
                 (ICPP 2019)},
    year = {2019},
}

HPAS is licensed under the BSD 3-Clause license.

[1] Emre Ates, Yijia Zhang, Burak Aksar, Jim Brandt, Vitus J. Leung, Manuel Egele, and Ayse K. Coskun. HPAS: An HPC Performance Anomaly Suite for Reproducing Performance Variations. In International Conference on Parallel Processing (ICPP), Aug. 2019

[2] Ozan Tuncer, Emre Ates, Yijia Zhang, Ata Turk, Jim Brandt, Vitus J. Leung, Manuel Egele, and Ayse K. Coskun. Online Diagnosis of Performance Variation in HPC Systems Using Machine Learning, in IEEE Transactions on Parallel and Distributed Systems (TPDS), vol. 30, no. 4, pp. 883-896, April 2019.

[3] Ozan Tuncer, Emre Ates, Yijia Zhang, Ata Turk, Jim Brandt, Vitus Leung, Manuel Egele, and Ayse K. Coskun. Diagnosing Performance Variations in HPC Applications using Machine Learning. In International Supercomputing Conference, ISC-HPC 2017., pp. 355-373, June 2017. Gauss Award.