Nehalem Performance Preview

April 7, 2009 by

Pages: 1 2 3 4 5 6 7 8 9 10 11 12

Myrimatch

Myrimatch is a tool that identifies proteins by comparing the mass spectrometry data for each constituent peptide against a known database of proteins. The application was developed by Dr. Tabb at Vanderbilt University, and is both multi-threaded and can be clustered efficiently across multiple systems. The data for the benchmark is a set of experimental spectra data collected from baker’s yeast and compared against a FASTA database from yeastgenome.org. Thanks go to Scott Wasson of Tech Report for sharing this benchmark with us. Like Euler 3D, the number of threads can be varied, so all our results are presented over a range of thread counts.
Figure 13 – Myrimatch Performance
Myrimatch performance is reported in execution time, and here Nehalem shows a limited (compared to Euler 3D) but respectable 30% advantage over Harpertown at peak thread count, and a smaller 20% gain with a single thread. The performance was quite stable thankfully with no variation at a given thread level. However, the scaling still poses a bit of a puzzle, as performance barely improved (only 30%) from 4 to 8 threads, while going from 8 to 16 threads increased performance by 90%. In general, scaling should almost always fall off, rather than increasing with thread or processor count.
Figure 14 – Myrimatch Power Consumption
The figure above shows power consumption for Myrimatch at different thread counts; at the highest thread level, Nehalem consumes about 60W less than the Harpertown system. Yet again, the Windows scheduler appears – notice the ~24W jump in the middle of the 8 thread run on Nehalem. That’s a thread being migrated from a virtual CPU to a physical one; it could be that SMT really doesn’t add a whole lot here and that’s why the scaling from 4 to 8 threads was so poor for Nehalem.
Figure 15 – Myrimatch Energy Efficiency
Since we used execution time above, we will use energy consumption as our metric here. The graph doesn’t do a great job of showing it, but at the highest thread counts, the Nehalem system uses about half as much energy as Harpertown (~6500 joules vs. ~11500).

Pages: « Prev   1 2 3 4 5 6 7 8 9 10 11 12   Next »

Discuss (52 comments)