Data Staging/Measurements

Measurements

Measurements on 20.07.2010

Purpose: to test how the introduction of DTRList class, based on C++ lists, would impact the performance of the scheduler

Description: 10000 DTRs were created in the generator, every of them was assigned a random status. The scheduler was changed to wait until all the 10000 were gathered in the DtrList and start the procession only after this. This program was run for 15 times, collecting the results every time.

Results:

• The procession (the time until all the 10000 DTRs were returned back to the generator) took 10 seconds in average
• The procession required about 2400 iterations of the main scheduler loop
• The creation and copying of 10000 DTRs took in average 1m 10 sec
• CPU usage was 93% in average during creation and copying of DTRs, around 65% during the procession
• Maximum memory usage was 330m virtual, 54m physical

Measurements of effects of Credential member in DTR (16.08.2010)

10000 DTRs were created and the memory used measured

With the Credential member:

• 360m virtual, 346m physical

Without:

• 47m virtual, 32m physical

This factor 10 increase can be avoided by not storing the whole Credential, but rather the information that we need from it (which is all static), such as DN, expiry time etc as simple strings/ints.

The following table shows the physical memory increase over time, showing it is roughly linear

while(true); do ps avx | grep DTRMemTest | grep -v grep | perl -ane 'print " $F[3] $F[7]\n";'; sleep 1; done
0:00 12172
0:01 18320
0:02 24556
0:03 30684
0:04 36840
0:05 42884
0:06 48852
0:07 54788
0:08 60548
0:09 66304
0:10 72332
0:11 78044
0:12 84176
0:13 90300
0:14 96428
0:15 102368
0:16 108576
0:18 114624
0:19 120404
0:20 126528
0:21 132660
0:22 138868
0:23 145048
0:24 151224
0:25 157432
0:26 163636
0:27 169796
0:28 175924
0:29 182040
0:30 188152
0:31 194148
0:32 200248
0:33 206404
0:34 212528
0:35 218684
0:36 224888
0:37 231040
0:38 237192
0:39 243268
0:40 249420
0:41 255652
0:42 261836
0:43 268040
0:44 273952