Online Level-3: Performance Filter Standalone

Performance Test Filter Standalone

Timing studies for regional tracking using a Pentium II, 400MHz system running Fermi Linux 6.1.

• Efficiencies of unoptimized vs optimized code are different. Shocking!
• The problem is with COT tracking, the Silicon stand-alone seems fine. Thus in one tracking test with optimized code COT tracking found 317 tracks and with unoptimized 575.
• No serious effort has been made so far to make the algorithms faster. Improvement expected.

So in summary, the numbers which should improve are given below, the Silcon numbers are most reliable, COT tracking seems to be the culprit in the optimized mode. COT tracks are seeded from SVT track. SVX tracking from preceeding COT track. The numbers are calculated per track and should not depend on the above described problem with difference in optimized and unoptimized code.

Algorithm Studied	Unoptimized	Optimized
Regional COT	0.02sec	0.01sec
Regional SVX	0.19sec	0.14sec
Vertex Constraint	0.04sec	<0.04sec

Assuming that optimized code is debugged in the coming months for a simple two track trigger with regional tracking around SVT tracks followed by a vertex constraint the time should be 2x(0.01+0.14)+0.04 < 0.34 seconds. For unoptimized code this would be 2x(0.02+0.19)+0.04 = 0.46 seconds. Needless to say the next step is to speed up existing algorithms and to regionalize existing faster algorithms like outside in (OI) tracking.

As a corollary, based on tests done by Steve Timm in FCC, we might expect a performance improvement of a factor of 2.0 per CPU between this system, and the system we intend to buy for Run II. The latter is most likely going to be a 800MHz Pentium III Dual processor system.

Less than 0.17 seconds per event ought to be well within reach for the 2-track multi-body hadronic trigger stream and even allow more fancy algorithms.