Commit 218edc0c authored by Jonathan Jenkins's avatar Jonathan Jenkins
Browse files

progress on 0.5.0 release notes

some TODOs still - see "___" placeholders.
parent 0efafcdc
codes-base and codes-net have been combined into a single project (now at
updated to ROSS revision ___
fixed a large number of warnings across the codebase.
addition of the SlimFly network topology, corresponding to the Wolfe et al.
paper "Modeling a Million-node Slim Fly Network using Parallel Discrete-event
Simulation", to appear at SIGSIM-PADS'16. See ___
modelnet now supports sampling at regular intervals. Dragonfly LPs can
currently make use of this - others can be added based on demand. See
allow 0-byte messages in model-net.
enable "local" model-net messages (for LPs sharing the same model-net endpoint -
approximates a zero-copy ownership pass of the payload)
the model_net_event family of functions now return a token value which must
be passed to model_net_event_rc2.
concurrent workload support added to workload generators and the MPI
simulation layer. See codes-jobmap.h and the modified codes-workload.h.
Thanks to Xu Yang for the partial contributions. Not all workload generators
support concurrent workloads at this time.
scripts for generating job allocations specific to the torus and dragonfly
topologies. See scripts/allocation_gen. Thanks to Xu Yang for the
multiple fixes to the MPI simulation layer and DUMPI workload generator.
concurrent workload support and more flexible rank mappings in the MPI
simulation layer. Thanks to Xu Yang for the initial code.
removed scalatrace workload generator, which never made it to a usable state.
a new checkpoint IO workload generator has been added, based on the Daly paper
"A higher order estimate of the optimum checkpoint interval for restart
dumps" at Future Generation Computing Systems 2004. See ___.
fixes to rc-stack - memory no longer leaks in sequential mode, and optimistic
debug mode is now supported.
added a more performance-sensitive function, codes_mapping_get_lp_info2, to
codes-mapping (passes const pointers around instead of copying strings)
added a "mapping context" API for better controlling implicit LP-LP mappings,
including modelnet, local storage model, and resource. See
codes/codes-mapping-context.h and added functions in the mentioned LPs for
formalized a callback mechanism for CODES, replacing previous ad-hoc
methods of passing control from LPs to arbitrary other LPs. The request and
local storage model LP APIs have been changed to use this mechanism, and
model-net has additional APIs to use user-provided mapping contexts. See
codes/codes-callback.h for the API and tests/resource-test.c for advanced
model_net_event_rc (use model_net_event_rc2, which will eventually be renamed
to model_net_event_rc)
codes_event_new (define your own bounds-checking macro if need be)
significant source reorganization / refactoring
refactor some private headers out of the public eye
dead code removal
improved example_heterogeneous example program
added configuration to example_heterogeneous showing two torus networks
reorganized files to prevent name collisions on OSX. Top-level docs other than
copyright now in doc directory
additions to best practice document
more stable file format for configurator output
ignore unrelated parameters passed into filter_configs
handle empty cfields in configurator
combined network and IO workload APIs into a single one
adding dumpi workload support in codes-workload-dump utility
workload dump utility option cleanup
renamed "bgp" workload generator to "iolang", significant cleanups
put network workload ops in workload dump util
removing one of the dumpi libraries from the build. It was generating some unwanted dumpi files.
network workload API more fleshed out
configuration bug fixes for larger LP type counts
resource LP annotation mapping hooks
local storage model API switch to use annotations
better configuration error handling
hedge against precision loss in codes_local_latency (see codes.h)
use a different RNG than default for codes_local_latency
- prevents addition/removal of codes_local_latency calls from poisoning RNG
stream of calling model
added simple GVT-aware stack with garbage collection (see rc-stack.h)
cleanup of much of the code base
more informative error for failure to find modelnet lps
removed redundant include directory on install (was 'install/codes/codes/*.h')
reorganized files to prevent name collisions on OSX. Top-level docs other than
copyright now in doc directory
updated code documentation
fix linker error in certain cases with codes-base
tweaked config error handling
fix to loggp latency calculation when using "receive queue"
made torus lps agnostic to groups and aware of annotations
miscellaneous fixes to dragonfly model
updates to simplep2p: support for having different latency/bw at sender &
receiver end. See src/models/networks/model-net/doc/README.simplep2p.txt
minor fixes to usage of quickhash in replay tool
fixed RNG reverse computation bug in loggp
fixed swapped arguments in round-robin scheduler causing short circuit
minor changes to dumpi trace config files
resolving minor bug with reverse computation in dumpi traces
Updating network trace code to use the combined workload API
Adding synthetic traffic patterns (currently with dragonfly model)
Adding network workload test program for debugging
Updating MPI wait/wait_all code in replay tool
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment