GitLab

Somehow the bug where the first destination of the random permutation
pattern wasn't randomized (but instead defaulted to rank 0) returned.
This resulted in attempted self messages if terminal 0 is allocated
for the random permutation workload as it will, by default, attempt
to send messages to itself resulting in a failed assert.

This fix checks to see if the rank has generated any data yet. If
it hasn't, then it will pick a non-self random destination and will
send to that destination until the random permutation threshold has
been met.

During the creation of the dragonfly-dally.C model, QoS features
were added. The addition of these features for some reason
also removed lines handling last_buf_full behavior. This resulted
in busy time always being reported as zero. I've added lines from
dragonfly-custom.C regarding this to restore previous behavior.

There are still added lines from QoS regarding last_buf_full and
busy time in router_packet_send() that are mildly confusing but
don't appear to affect final results and are left as is.

ROSS updates

See merge request !82

The feature addition of the compute_time_speedup factor did affect
the compute time correctly. It, however, didn't affect the scheduling
of the new event after the compute time. This resulted in the delay
savings from the compute time speedup being applied to the comm. time
instead. This has now been resolved.

Core CODES Updates

See merge request !85

This commit adds functionality to accept a compute_time_speedup="X"
parameter in a model's configuration file to accelerate the delay
applied when accounting for compute_time in model-net-mpi-replay.c.

X here is a double where 2.0 represents a 2x reduction in delay
resulting from compute_time simulation. e.g. a setting of "2.0" means
that all increments of compute_time are reduced by a factor of 2.

Update all usage of "server" lp to "nw-lp"

This change is made for consistency across all configuration files
and workload generators in core CODES. Where older models and versions
of CODES utilized the "server" identifier to specify the number of
server LPs per modelnet group repetition, newer models and
configurations utilized the less paradigm specific "nw-lp" identifier.

This shift slightly fractured the codebase and led to confusion when
running similar/same configurations across different network models.

Specifically, this change converts all usage of the "server" identifier
to the "nw-lp" identifier. All tests, existing default configuration
files, and models have been updated to use the "nw-lp" identifier for
the workload server LPs. Some docs and readmes have been updated as well
to no longer instruct new users to use the "server" identifier.

Revert "Merge branch 'workloads' into 'master'"

See merge request !78

This reverts merge request !77

Workloads

See merge request !77

# Conflicts:
#   maint/codes.pc.in

Adding default definition to pkg-config variable: swm_datarootdir

Closes #167

See merge request !76

set eager threshold to max for conc-online workload type, since rendezvous protocol cannot work with Conceptual for now

DFP Script Fixes

See merge request !66

ross-org url (and whitespace)

See merge request !74

Replace call of deprecated MPI_Type_hindexed

See merge request !73

Fixing multiple ranks per node mapping selection

See merge request !71

Compatibility fix of merged QoS features

See merge request !72

The commit replaces the call of the deprecated function MPI_Type_hindexed
with the current function MPI_Type_create_hindexed.

Using CODES with OpenMPI depends on this change.