| Commit message (Collapse) | Author | Age |
| |
|
| |
|
| |
|
| |
|
|
|
|
| |
These haven't been used since Joshua's PRP.
|
|
|
|
|
| |
- Fix floating point numbers in L3 allocations
- Fix truncation of texture_synthesis' name
|
|
|
|
|
| |
Printing is too slow for many Level-A and -B tasks, so we limit it
to Level-C.
|
| |
|
| |
|
|
|
|
|
|
|
| |
- Fix leaks in localization, multi_ncut, svm, and texture_synthesis
- Lower preallocation for sift and texture_synthesis to match memory
profile with QCIF
- Remove trailing whitespace in script_localization.c
|
|
|
|
|
|
| |
We were leaking the results and some intermediate state every job.
Also updates the size of the preallocated memory to 16MB (should be
plenty, as the WSS per Massif is only ~1MB for qcif).
|
|
|
|
|
|
|
|
|
| |
- Add a unified runner script for DIS
- Remove explicit case study termination script
(do `kill ${cat pids.txt}` instead)
- Auto-create bin directories for TACLe
- Add the python cache directory to the gitignore
- Tweak input path representation in benchmark name files
|
|
|
|
|
|
|
| |
- Quote lookup ID in ID2PID
- Fix SD-VBS benchmark names
- Use bash rather than sh as the shell so that `echo -e` works
- Remove dispatch limit that was for debugging
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
New features:
- Checks arguments
- Checks that we're root first
- Checks if task launches succeed
- Checks that a case study is not already running
Fixes:
- Removes invalid uses of sudo
- Fixes use of ',' rather than ';' in resctrl specification
- Fixes PID determination
- Always launches Level-C tasks as asynchronous threads
Additionally converts the whole file to only use spaces and to be
directly executable.
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
|
| |
- Fixes SD-VBS build with LITMUS 1 defined
- Fixes memory leak in stitch and tracking benchmarks
- Adds initial case-study script and configurations
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Adds <period> and <criticality level> parameters to benchmarks and
disables cache flushing when compiled with LITMUS. Also fixes build
with LITMUS enabled on yamaha.cs.unc.edu.
Also allows output to be disabled, automatically infers _rt_lock_id,
and allows for a specific pair ID for paired tasks. (The pair ID is
added to all the semaphore and shared memory names, allowing multiple
pairs to run in the system at a time.)
|
|
|
|
|
|
|
|
| |
Remember, M_i values are the worst-case slowdown of the average-
-case execution time.
Also factor out functions shared with M_ij computation code to a
shared library.
|
|
|
|
|
|
|
| |
- Add a run_all_sd-vbs.sh script
- Fix a typo in run_all_tacle.sh
- Fix an issue with reversed arguments for asynchronous task pairs
in run_bench.sh
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
SD-VBS:
- Run silently
- Fix some whitespace errors
- Don't duplicate extra.h
- Auto-detect if building with LITMUS-RT
- Disable result checking
- Add helper symlinks
Misc:
- Remove unused code from libextra
- Set some missing rt_param fields in libextra
- Disable CSV info dump from computeSMTslowdown.sh
- Widen scope of .gitignore on .txt files
- Include list of 2MB DIS pair benchmarks and inputs
|
|
|
|
|
|
|
|
|
|
|
| |
Includes the SD-VBS benchmarks modified to:
- Use libextra to loop as realtime jobs
- Preallocate memory before starting their main computation
- Accept input via stdin instead of via argc
Does not include the SD-VBS matlab code.
Fixes libextra execution in LITMUS^RT.
|
| |
|
| |
|
|
|
|
|
|
| |
- anagram: Fix two off-by-ones, the heap size, and the result
- rijndael_dec: Reset result checksum at job start
- huff_enc: Reset mutable global variables at job start
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This allows `run_bench.sh` to replace all previous general paired
benchmarking scripts. See `run_everything.sh` for an example of how
to use the new script.
This also fixes two critical bugs present in the old scripts:
1. The cpus_list was incorrectly set for resctrl when running
synchronous pair timing with L2+L3 isolation (i). Each write
to that file replaces the previous contents, and it was previously
written once for each core. This prevented the entire 1st core
from using any L3 cache.
2. The competing task list was indexed improperly in the asynchronous
pair timing script. This caused the competing task to always be
identical to the task being profiled.
These issues combine to imply that in the rejected RTSS'20 paper:
- All Level-C SMT timing data is suspect
- Level-A and -B SMT timing data with L2+L3 isolation (i) is
wildly optimistic
|
| |
|
| |
|
|
|
|
|
| |
Now supports specifying an input generation command for
each benchmark - see dis/dis2MbInNames.txt for an example.
|
|
|
|
|
| |
Also move run_all_dis.sh up a directory in preperation for
further changes.
|
|
|
|
|
|
| |
- anagram, audiobeam: Reset heap in _init() function
- g723_enc: Don't use static local variables
- huff_dec: Reset all global state in _init() function
|
|
|
|
| |
No other changes.
|
|
|
|
|
|
|
|
|
|
| |
- Use consistent approach to prevent computations from being
optimized out in the DIS benchmarks
- Don't print results in the DIS benchmarks
- Fix a memory corruption in TACLe's "epic" benchmark reflect1()
- Fix return codes for all benchmarks
- Rename run_baseline.sh to run_bench.sh in preperation for this
being the main benchmarking script
|
|
|
|
|
|
|
|
|
|
|
|
| |
Changes to DIS code:
- field, pointer, transitive, and update's random initialization
steps moved the main job loop (so that they run on fresh state
each job).
- Moved free() calls outside of the job loop in matrix
- Removed loose clock() call in pointer
Misc fixes:
- Added input file for neighborhood
- Log status before, rather than after, printing in gen_input.py
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There was previously a huge amount of shared code that had to be
copied back and forth. This should reduce the maintenance burden
by containing all future changes to a single file.
New unified library is fully backwards-compatible but also
introduces and the easy-to-use `for_each_job` macro which
replaces the specific `for(...) START_LOOP ... STOP_LOOP`
format requirement and is generally much harder to abuse.
New unified library also automatically cleans up its shared memory
and semaphores, so this commit also removes the separate
`cleanupSemaphores` binary.
I also found a precursor of `extra.h` written by Sims in
`litmusStuff.h`. This code is only interesting for historical
purposes, so it is also removed in this commit.
This commit also adds debug options to all the Makefiles and
silences rm's complaints about non-existent files in make clean.
|
|
|
|
|
| |
New macro `for_each_job` simplifies instrumentation.
Backwards-compatible.
|
|
|
|
|
| |
Everything without isolation: run_tacle_rtss20.sh
For isolation you have to run the hodgepodge of scripts as before.
|
|
|
|
|
|
|
|
|
| |
DANGER! There's an inconsistency in the input files used for the
all-pairs tests and those used for the baseline tests.
Baselines: run_all_dis.sh calls run_dis.sh to do all baselines.
All-pairs: Uses old TACLe script (not included here).
Cache allocation 2nd-order effects on SMT friendliness: run_pair_convexity.sh
|
|
|
|
|
| |
Also fixes a missing header in DISstressmarkRNG.h and the "all"
make target. No functional changes.
|