Messages in this thread | | | From | (Larry McVoy) | Subject | Re: Interesting scheduling times - NOT | Date | Thu, 24 Sep 1998 16:49:32 -0700 |
| |
: I only mentioned that because Larry mentioned normal distributions of : results in a previous message and it struck me as strange to expect a : normal distribution of scheduler latencies. He was arguing against using : the minimum time from a benchmark, as I recall. There are a number of : places this is useful number, for instance, the round trip times from : ping. The minimum rtt is a rather hard limit and tells you what the best : latency you can expect is.
Well, this is a fine theory - and it was what I thought when I started measuring things - but it's wrong in practice. If you take a suite of tests, lmbench for example, and do a bunch of runs and scatter plot them and stare at them you'll see patterns emerging. Now if the pattern was that most run times clustered around the min, then my feeling is that the min is the right number. Wherever they cluster up is the number I wanted because that was the number mostly likely to be seen.
Note that if the min is where things cluster, then the min will be very darn close to the median.
The min is not necessarily the right answer. Suppose most of the time you get get N but one time you happen to get very lucky in your cache layout (or whatever) and you get N/2. Is N the right answer or is N/2? In my mind, the right answer is the one that is most frequently reproduced.
So, yes, you can get better numbers than what I report. But not with any high probability. Choosing the median was on purpose, I wanted numbers that represented what I thought people would actually get when doing similar things - I'm really sick of benchmarks that tweak things like crazy (i.e., -lspec_malloc) to get better numbers, numbers which are hard or impossible to reproduce in real life.
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |