This shows you the differences between two versions of the page.
Next revision | Previous revision | ||
difx:difx2.0baselbased [2009/10/29 07:37] adamdeller created |
difx:difx2.0baselbased [2015/10/21 10:08] (current) |
||
---|---|---|---|
Line 1: | Line 1: | ||
+ | |||
====== Baseline-based efficiency improvements in DiFX2.0 ====== | ====== Baseline-based efficiency improvements in DiFX2.0 ====== | ||
Line 14: | Line 15: | ||
The following diagram illustrates what I mean: | The following diagram illustrates what I mean: | ||
- | {{baselinebased_after.png}} | + | {{baselinebased_after.png?800}} |
Say you buffer 10 FFTs at each station. That means you only get a cache miss for 1 of every 10 cycles, rather than every single cycle. This is little enough to make the cache missing a relatively small addition to the total processing increase. (There are unavoidable non-linear increases in processing cost going to either more stations or more channels. With more stations, you have more cross-multiplications to do, since there are more baselines. This starts to get significant at around 10-15 stations. With more channels, the cost of the FFT is going up with log N. Since the FFT is now dominating in DiFX2.0, that means going from 128 channels to 2048 [x16, = x4 cost to FFT] is likely to hit you pretty substantially - a factor of 2-3 at least in station-based cost.) | Say you buffer 10 FFTs at each station. That means you only get a cache miss for 1 of every 10 cycles, rather than every single cycle. This is little enough to make the cache missing a relatively small addition to the total processing increase. (There are unavoidable non-linear increases in processing cost going to either more stations or more channels. With more stations, you have more cross-multiplications to do, since there are more baselines. This starts to get significant at around 10-15 stations. With more channels, the cost of the FFT is going up with log N. Since the FFT is now dominating in DiFX2.0, that means going from 128 channels to 2048 [x16, = x4 cost to FFT] is likely to hit you pretty substantially - a factor of 2-3 at least in station-based cost.) |