ecp-veloc / redset Goto Github PK

I've packaged redset for the gentoo overlay guru but our CI failed at testing it
https://bugs.gentoo.org/784668

Build log: https://784668.bugs.gentoo.org/attachment.cgi?id=701421
Test log: https://784668.bugs.gentoo.org/attachment.cgi?id=701424

REDSET 0.0.5 ABORT: rank 1 on localhost: Number of partner replicas (1) must be in range [1,1) with 1 ranks in set @ /var/tmp/portage/sys-cluster/redset-0.0.5/work/redset-0.0.5/src/redset_partner.c:116

Both XOR and PARTNER require at least two processes in each redundancy group. Currently the code segfaults if the user gives a failure group such that some process ends up in a redundancy group of size 1. To be helpful, we should print and return an error instead. With SCR, we would automatically fallback to SINGLE in such cases. That could be another option.

See SCR v3.0rc2 project: https://github.com/LLNL/scr/projects/3

Process

• Resolve necessary issues
• Merge any existing PRs
• Create a commit(s) which updates the project version number in the code and updates references to any dependency versions.
• Merge the commit
• Tag a release on GitHub
• Add release note (can be done after the release is tagged)
• Update the spack package, both with the version hash and the dependency versions.

Tests seem to cause the issue:

/tmp/veloc/redset/test/test_config.c:66:35: error: ‘REDSET_KEY_CONFIG_DEBUG’ undeclared (first use in this function); did you mean ‘REDSET_KEY_CONFIG_GROUP’?
   66 |   if (kvtree_util_get_int(config, REDSET_KEY_CONFIG_DEBUG, &cfg_debug) !=
      |                                   ^~~~~~~~~~~~~~~~~~~~~~~

In

the test assume that

${MPIEXEC} ${MPIEXEC_NUMPROC_FLAG} 2 -N2 ./redset_test

is generically valid, however -N is not a documented option of mpiexec and seems to mostly be an srun or so option to specify how many ranks per node to use. It fails eg if RESOURCE_MANAGER=NONE is used and OpenMPI is the mpi stack.

Process

• Resolve necessary issues
• Merge any existing PRs
• Create a commit(s) which updates the project version number in the code, CMakeLists.txt, and updates references to any dependency versions.
• Merge the commit
• Tag a release on GitHub
• Add release note (can be done after the release is tagged)
• Update the spack package, both with the version hash and the dependency versions.

See SCR v3.0 project: https://github.com/LLNL/scr/projects/4

Process

• Resolve necessary issues
• Merge any existing PRs
• Create a commit(s) which updates the project version number in the code and updates references to any dependency versions.
• Merge the commit
• Tag a release on GitHub
• Add release note (can be done after the release is tagged)
• Update the spack package, both with the version hash and the dependency versions.

More tests to be added to redset:

• Modify test so that it can run on a single host (avoid using hostname to split into redundancy groups).
• Add a test where different ranks write a different number of files, and include a case where one rank writes 0 files.
• Assign different groups to different files with chown(), and check that proper group is restored after rebuilding a deleted file.
• Create a test in which COMM_WORLD is divided into multiple redundancy sets. Delete files from one set so that it can't be recovered. Verify that all ranks in COMM_WORLD identify that the rebuild failed.
• Test that serial rebuilds work for all redundancy schemes.
• Run valgrind to check for memory leaks.

Resolve issues noted in:

• #20
• #30

See SCR v3.0 project: https://github.com/llnl/scr/projects/2

Process

• Merge any existing PRs
• Create a commit(s) which updates the project version number in the code and updates references to any dependency versions.
• Merge the commit
• Tag a release on GitHub
• Add release note (can be done after the release is tagged)
• Update the spack package, both with the version hash and the dependency versions.

Intel has an open-source library supporting vectorized functions to compute RS GF(2^8) math:

https://01.org/intel%C2%AE-storage-acceleration-library-open-source-version
https://github.com/intel/isa-l

This is licensed under BSD-3. It looks like it may also support processors other than just Intel, including PowerPC. This could offer a performance boost for our internal RS encoding.

For example, the gf_vect_mul function here:

https://github.com/intel/isa-l/blob/master/include/gf_vect_mul.h

might plug in for:

This will take some developer time to make the switch, and it adds a dependency. Also, if we are bottlenecked by storage or network bandwidth, reducing encode/decode time may not improve overall performance by much. So some work should be done to investigate before taking this on.

Looks like my Mac is picky about including HOST_NAME_MAX (see https://stackoverflow.com/questions/30084116/host-name-max-undefined-after-include-limits-h). It needs to be fixed in the CMake.