Development Cage Results
This page contains some preliminary results using our noise generation infrastructure on a 32 node develpment cage
Selfish
Selfish is a benchmark that measures the number of cycles of an application lost executing other code. In the graphs below a spike indicates a cycle loss was detected by the benchmark with the quantity of time lost indicated by the amplitude of the spike.
Linux
Linux Laptop
Linux Service Node
Catamount
Un-Modified
Noise Infrastructure but No Noise
10Hz, 0us Noise
10Hz, 10us Noise
10Hz, 100us Noise
10Hz, 1000us Noise
100Hz, 0us Noise
100Hz, 10us Noise
100Hz, 100us Noise
100Hz, 1000us Noise
1000Hz, 0us Noise
1000Hz, 10us Noise
1000Hz, 100us Noise
1000Hz, 1000us Noise
In this case we start missing timer interrupts so our efective interrupt frequency is 500Hz, 10000us
NAS Parallel Benchmarks
The following graphs show the performance of the NAS Parallel Banchmarks on Catamount/Red Storm.
Increase in Timer Interrupt Frequency Only
The difference in small (< a percent) but seems to be increasing and may be more pronounced at higher node counts
CG
I am rerunning to get the 16 node data as a compiler issue caused a problem in the previous
BT
EP
FT
IS
IS seems to be performing very slightly better at higher interrupt frequencies
LU
MG
SP
HPLinpack
5 x 5, N = 20000, NB = 160
No Noise
| Timer Frequency (Hz.) | Gflops |
|---|---|
| 10 | 66.16 |
| 100 | 66.14 |
| 1000 | 66.03 |
Noise
| Frequency: 10Hz | Frequency: 100Hz | Frequency: 1000Hz | |
|---|---|---|---|
| Delta: 0 us | 66.03 GFLOPS | 66.03 GFLOPS | 66.04 GFLOPS |
| Delta: 10 us | 66.03 GFLOPS | 65.88 GFLOPS | 65.38 GFLOPS |
| Delta: 100 us | 65.90 GFLOPS | 65.30 GFLOPS | 59.65 GFLOPS |
| Delta: 1000 us | 63.64 GFLOPS | 57.82 GFLOPS | 27.95 GFLOPS |
HPC Challange
The following are performance numbers of the HPCC benchmark running with my noise infrastructure on RedStorm/Catamount
COMMING SOON
