[LITMUS^RT] Schedule and Locking Overhead information

[Björn Brandenburg]

> On 11. May 2019, at 18:19, Ricardo Teixeira <ricardo.btxr at gmail.com> wrote:
> 
> I executed 800 experiments for an out of stock protocol, each experiment runned for 30 seconds. During the experiments, I got overhead information using the scripts disponible at github (throught the feather trace tools). 
> 
> I also ran the same set of experiments for another out of stock protocol and got the same overhead information. Both protocols generates nearly the same amount of jobs. 
> 
> Analysing the collected data, I realized that the second protocol generated about 3 times more samples than the first, for all types of data collected, for example: CXS, LOCK, RELEASE-LATENCY, RELEASE, SCHED and SCHED2. 

Hi Ricardo,

some variation in the number of scheduling decisions and context switches can be expected, especially if you are comparing different scheduling and/or locking policies, but release-related events are determined solely by the workloads, so if you are running the same workload under both setups, it shouldn’t differ by a factor of three. 

> 
> Another problem was that both protocols gererated a reasonable amount of LOCK samples, but an insignificant amount of UNLOCK samples. 

That’s very strange — the operations should obviously be symmetric and hence produce an equal number of samples. 

> For example: for an subset of 100 experiments, there was 62k LOCK samples for the first protocol and 177k for the second, there was 0 UNLOCK samples for the first protocol and 6 for the second. That situation was present in all of the 8 subsets, where the amount of UNLOCK samples was never above 10 for both protocols. 
> 
> Through the tracers, I known that the LOCK and UNLOCK operations were happening as expected and I not changed the code which saves the overhead tracers. 
> 
> Could anyone help me to understand what is happening? 

Did you see any messages about failed writes to the trace buffers? 

- Björn

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 5061 bytes
Desc: not available
URL: <http://lists.litmus-rt.org/pipermail/litmus-dev/attachments/20190512/f5abda71/attachment.bin>