[Barrelfish-users] Question about running ump_bench
reto.achermann at inf.ethz.ch
Mon Jun 24 10:14:57 CEST 2019
Apologies for the delay.
That's interesting. To recap, when you set MAX_COUNT > 4096 then the latency
gets smaller. The size of the result array should not have an effect on the
measurements, I suppose.
How do you do the benchmark with just one timestamp instead of the array and
what value do you obtain there 800 or 500?
Try to run with 6000 iterations where you get the lower latency, and first
ignore 2000 iterations and store 4000 timestamps. How does this turn out?
On Tue, 2019-04-30 at 16:38 +0800, 李康宇 wrote:
> Dear all,
> I'm trying to run /usr/bench/ump_bench on zynqmp ZCU104 in order to figure out
> the time taken by ump communication. However,I found that the time results
> will change as the number of experiments changes.
> Take /usr/bench/ump_bench/latency as an example,when I change the constant
> "MAX_COUNT" to 4000,the time results distribute around 800(cycles),and when I
> set "MAX_COUNT" to 6000,the time results distribute around 500.I also found
> the point of change is at 4096,in other words,there are two results when
> "MAX_COUNT" is greater than or less than 4096.I noticed that the program would
> first apply for memory space of sizeof(struct timestamps) * MAX_COUNT. I want
> to know if this affects the results of the experiment, because when I use only
> one timestamp instead of a timestamp array, the experimental results are
> consistent.If so,I also want to know why and how this could affect the time
> Barrelfish-users mailing list
> Barrelfish-users at lists.inf.ethz.ch
More information about the Barrelfish-users