Tuning and Analysis Utilities

TAU

PRL

[PARP Logo] The Parallel Recursion Project(University of Oregon)

The PaRP (Parallel Recursion Project) is a software package for materials science numerical simulations on high-performance computing platforms. It is based on the dynamic recursion method, which is a method suited for physical systems that can be expressed in terms of sparse matrices (Hamiltonians).

[Parp

The above figure illustrates the use of TAU User Defined Events for tracking memory allocation in multi-threaded programs. The user event data is maintained on a per thread basis and TAU maintains statistics such as standard deviation, mean value etc. of the samples collected.

[Parp

TAU was used to analyse the time spent in different locks in Parp. Parp uses pthreads for parallelism and the timing data is displayed on a node, context and thread basis. For more detailed information, hardware performance counters can be used to track counters such as the number of secondary data cache misses in each profiled block per thread.

The details of this run are: 


# Diagnostics information for PID: 13691, Anderson transistion (W = 0) dynamic recursion, 50 recursion levels
# Clustersize: 64, threads: 4 (12 procs), partition size: 4096 (level 1)
#Pruning threshold below max:	0
# Physical memory configured:	767 MB
# Columns: recursion level; unit cell instantiations/iteration; branch object instantiations/level;
# length of branch addition list/level; number of activated clusters
Lvl: 49, part: 123, act nodes: 259840, act cells: 4060, br's: 844 act.ratio:64
Execution Time:
Total (process level):		000:04:48.873
Thread 0:  1144 jobs
Thread 1:  1103 jobs
Thread 2:  1075 jobs
Thread 3:  1058 jobs
All threads: 4380 total jobs
Top level in tree: 6
Global maximum of yn value:	0.0582386
Total wait time % of total run time: 0%

System time used: 000:00:05.843, user time used: 000:09:28.128
Max. resident memory size used: 9968 KB
Minor page faults: 140, major page faults: 0, swap outs: 0
Voluntary context switches: 25234, involuntary context switches: 5852
Total number of nodes generated: 259840

URL: http://www.csi.uoregon.edu/parp