Hack to force replay tool to use one preallocated and initialized
I/O buffer for all I/O operations. Malloc'ing a new
(uninitialized) buffer for each I/O operation causes really poor
performance, likely due to page faults.

Darshan generates longer delays, and mpi replay tool now can be
configured to play a percentage of the overall delay.

Update recorder workload generator to obtain workload operation
parameters from new Recorder I/O traces (courteousy of Babak
Behzad). I/O offsets are now tracked explicitly in the recorder
traces, and workload file_id's are set to original execution file
descriptors.

Darshan workload generator now evenly distributes aggregators
acoss the entire rank-space. Also, individual operation sizes
are evenly distributed among aggregators, so aggregator
workloads are more evenly distributed.

- currently ugly-looking, might want to think of
  a better way to acheive this

Delay calculation now estimates duration of barriers (collective
communication).

The collective buffering algorithm now appropriately assigns i/o
sizes and offsets according to an aggregator's "file view".

For a particular LP type:
- get the global number of LPs of that type
- get the relative ID w.r.t. LPs of the same type

Adding the codes-network-workloads API: currently there is support for reading in MPI traces generated through scala-trace. See network-workload/README for usage of test program.

The plugin was not destroying a hash table properly, causing
segfaults. Also, the workload did not close open files or
directories, causing an error on the number of open file
descriptors.

Darshan reports zero-valued timestamps as negative values, so
add check to set these timestamps to zero if they are negative.