Context: Options with values are not properly parsed.
If the value is not set, a bug makes appear that 0 was
provided.
This patch fixes options parsing (issue #337). An error
is reported if 'backend.option=' argument is provided.
* Basic support for memory allocation on GPU using CUDA unified memory. Partially addressing #284. IOR support completed.
* Support for GPU alloc in MDTest and MD-Workbench
* Option: support repeated parsing of same option (allows option sharing across modules).
* Checks for gpuDirect
* Integrate gpuDirect options and basic hooks, more testing to be done.
* POSIX: basic gpuDirect implementation working with fake-gpudirect library.
* CUDA allow setting of DeviceID for IOR (not yet MDTest).
* CUDA/GPUDirect Support --with-X=<path>
* Bugfix in option parser for flags that are part of an argument for an option, e.g., -O=1, if 1 is a flag it is wrongly assumed to be a flag.
* Improve and fix the reporting in MDTest. Reporting per rank now outputs the performance of the individual rank (before the barrier), the iteration throughput includes the barrier time (if not -B=0 is set). Before, it was the time after the barrier.
* Clarify the computation of the results.
* MDTest improve CSV output to include aggregated result.
* MDTest allow storing information per rank for later analysis when using the --saveRankPerformanceDetails=<FILE> option
* MDTest: refactored calculation of results, added time_before_executing a barrier.
Uploaded new mdtest patterns. Make them invariant to line nrs.
Added md-workbench tests.
Added --oversubscribe option if MPI understands it (needed for some systems)
IOR by default outputs the numbers of errors.
Reason: The amount of error messages can be overwhelming, particularly in a parallel program (Gigabytes...).
One -v increases the verbosity level to provide the extra details without adding too many other messages.
The SVCL argument is being removed from the DAOS API, so conditionally
compile out while maintaining backwards compatibility to versions where
it's still required.
The calculation per iteration first computes the value of the slowest process, i.e., highest time or lowest rate. This is then the value for the iteration.
Secondly, calculate the min/max/mean across iterations.
For tree operations, the value is identical to previous as only Rank 0 is involved.
* HDFS module ported to current AIORI API and extended
* Added instructions how to build and run with a HDFS.
* Replaced read with pread to support offsets.
* Implemented HDFS MDTest API
* Improved sync semantics
Julian Kunkel