- MPI 1.) Reporting MPI functions was fixed at 5. (Broke codes with less.) (Fixed) Changed to gathering top 20 or all greater than 0.5%. 2.) Long jobs broke because mpiP stats returned time in sci. notation. (Fixed) 3.) MPI fraction for mpiP now reports coefficient of variation. 4.) Changed reporting values of MAX (%) to Average. 5.) Fixed recent change to broke reporting memory in summary for multi-nodes runs
*NUMA 1.) Incorrectly stated 4MB pages instead of 2MB. (Fixed)
*GPU 1.) On dual system (nodes with/without GPU) report GPU is active, but not used on non-GPU nodes. 2.) No longer necessary to set REMORA_CUDA to get GPU information on gpu nodes. Can turn off by setting REMORA_CUD to any value except 1.
*MEM 1.) Summary now reports MAX VIRT, MAX PHYS and AVAIL at beginning of job.
-
numa
- Corrected 4MB to 2MB in description of pages.
-
remora_finalize.sh
- Create local nodes array, NODES_LIST, for accessing node and pid simultaneous as list (e.g. ${PID[$idx]} and ${NODES_LIST[$idx]} in loop)
-
install.sh
- config/modules must be changed from ib,NETWORK opa,NETWORK if OPA is detected
-
scripts/remora_finalize.sh
- Include impi fraction and breakdown links under "impi utilization".
- impi_fraction.html and impi_breakdown.html links in remora_summary.html were not
- included. Now, it is included if impi is found in REMORA_MODULES[@] in remora_finalize.sh.
- (This works similar to including the lustre_aggregated.html under "lustre utilization".
-
modules/numa
- numastats does not show THPs.
- Replace misses with foreign hits--since it is the real miss.
- Now uses vmstats to give Top-Right plot with allocated memory from thp and (thp+4k) pages. How to read text was placed at top.
-
modules/opa
- Fixed modules/opa.
- packs and bytes strings were never initialized. pack was used in string names instead of packs.
-
modules/network
- ibtracert not found in default path.
- Added /sbin
-
modules/ib
- was not set up to handle dev_hfi1_0 device for opa
-
docs
- added modules_help and modules_whatis files for supporting mkmod
Small issue fix (#45). Also ensures that there are extra checks when no scheduler is present.
Several issues fixed (#37, #40, #41, #42, #43) and partial fix for issue #39. For #39, the CPU collection is now correct when mpstats is not in the path of all nodes
Added MPI statistics collection for Intel MPI and Mvapich2 Modified how top level functions are called in order to improve overall handling Improved verbose mode information
Now collection and post-processing work for users with default csh/tsh shells.
Thanks to Mark Reed (UNC) for the suggested changes.
Added power and temperature modules. Added automatic discovery of NUMA nodes. Supports PBS scheduler. Supports running on a local machine (no scheduler needed). Multiple jobs for the same run: new folders are created. Use Google Charts to create plots automatically (no need for post-processing Python script except for network trace). Added a script to monitor free memory that kill the application before the OOM does. Added a script to generate post-processing files in case the node crashes.
Contributions from Jacob Pollack at Shell (Eth monitoring, CPU utilization post-processing, and 64-bit counters for Infiniband monitoring), Kevin Manalo (suggestion of automation of mlx device discovery) and wpoely86 (PBS support).
Added "eth" module which tracks Ethernet statistics in a similar fashion to how "ib" module tracks InfiniBand statistics. Added support for use of 64 bit InfiniBand counters if they exist. Added average CPU usage information to CPU plots.
Introduced real-time monitoring capabilities, enabled by REMORA_MODE=MONITOR. Fixed minor bug introduced in summary report warnings for systems with over 32 GB or memory per node. Updated user documentation.
Improved performance by allowing the use of a separate REMORA_TMPDIR which can be a local file system.
Simplified "remora" script itself by having a remora_init / remora_collect / remora_finalize scripts.
New "src/scripts" directory to clean source structure.
Cleaned up output directory and improved summary formatting.
Using "remora_env.txt" file in shared location REMORA_OUTDIR to capture remora options during runtime.
Added DVS file system data collection and file system blacklist.
Improved the format of the output files. The columns have a fixed width so that they are easier to read. Implemented a modular design, where all the data captured by REMORA is done by independent scripts in the modules folder. A configuration file, that needs to be configured as part of the installation process, specifies the active modules for a particular site.
Changed remora_post to work with older versions of Python. It was only working with Python 2.7+
Added memory monitoring for NVIDIA GPUs
Minor bugfixes
Fixed Lustre data parser
First public release of REMORA (REsource MOnitoring for Remote Applications) Remora is a tool to monitor runtime resource utilization:
- Memory
- CPU utilization
- Lustre usage
- NUMA memory
- Network topology
To use the tool, modify your batch script and include 'remora' before your script, executable, or MPI launcher.