[Planetlab-users] concerning the behavior of planetlab2.cs.virginia.edu in the last two weeks

[Jeff Hoye]

Livia,

We are running an ePost node on that computer, but I don't suspect it is a major consumer of CPU.

   However, may I suggest that execution time is not a good measurement of your application performance in PlanetLab?  If you look 
at a typical planetlab load average (using top) you will find that it is usually above 10 (I've seen it as high as 60 sometimes). 
  Interestingly (and obviously) these load averages go up dramatically before a major conference deadline.  There are probably a 
lot of applications running on any particular node, and we are all competing for CPU.  I would imagine that there are also all 
sorts of periodic tasks in experiments that will cause the load to fluctuate dramatically.  Because of all of this, I'm concerned 
that you will have way too much noise in your experiments if you use execution time as the metric.  Perhaps someone can suggest a 
system tool that will tell you how much CPU your program has used so that you can get a better idea of this metric.  As far as I 
know there is no system in PlanetLab to schedule CPU time on particular nodes, though there is a signup for bandwidth.

Also, I know that not everyone using planetlab reads every message on this list.  So you probably aren't going to get a complete 
answer as to who is running applications on a particular node.  Though I recall seeing a status page somewhere that says who is 
the primary user of any particular resource on a per-node basis.

All, let me know if I am incorrect on any of this.

Thanks!
-Jeff


Livia Maria Rodrigues Sampaio wrote:
> Hello all,
> 
> I have been running some experiments using a distributed application 
> (implemented by myself) in the planet-lab with 5 nodes, including 
> planetlab2.cs.virginia.edu. In the last two weeks the performance of my 
> application slowed down. As such performance depends on the workload of 
> the node planetlab2.cs.virginia.edu I suppose this node slowed down due 
> to some reason. I would like to have more information about such 
> workload variability because this is very important to my performance 
> studies. Particularly, If you have also been using this node in your 
> experiments you can contribute with my studies.  
> The objective of my experiments is to analyze the performance of a 
> distributed
> application using the execution time as performance metric.
> In one of the experiments I used
> messages of 32KB. The execution time of the application was around 1 
> second.
> However, from july 29th to august 13th, such execution times increased to
> 180 seconds. Presently, august 16th, the execution times are again around
> 1 second. The variability in the performance of the application is 
> associated with the workload of the node  planetlab2.cs.virginia.edu. It 
> seems that the node was "fast" and became "slow" for some
> period. This is not surprising in the planet-lab as pointed out in the 
> paper "Fixing the embarrassing slowness of OpenDHT on Planetlab". 
> However, I would like to have more information about the slowness of 
> this specific node.
> In order to put some light on this topic I started investigating two 
> possibilities: variabilities in the network load (to/from the
> node) or in the processing load (in the node). In the first
> case, I collected some
> traceroute data, in the periods the node was "fast" and "slow", trying 
> to find some
> communication bottleneck - in this case, I configured the traceroute to 
> send messages of 1KB. The traceroute data described the transmission of 
> a message from each of the 5 nodes considered in my experiments to the 
> same destination. I didn't noticed anything strange in the logs. 
> Consequently, I suppose the reason for the variability in the 
> performance of my application was due to an intensive use of the 
> resources (eg. CPU, memory) of planetlab2.cs.virginia.edu by a number of 
> applications running on this node, during the observed period, 
> concurrently with my own application. Particularly, memory bound 
> application can cause more swapout of the files been used. Moreover, 
> applications with high priority can use more cpu cycles than the others.
> 
> Another possibility is the intensive use of bandwidth causing 
> fairsharing of available bandwidth, in this case, each slice of the same 
> node cannot use the bandwidth limit specified for it. I suppose this is 
> more critical as the number of applications running in different slices 
> but on the same node increases.
> 
> I would appreciate a lot if you could help me on this topic in the 
> following way:
> 
> 1. Have anybody run applications on planetlab2.cs.virginia.edu during 
> the period from july 29th to august 13th? Is your application memory 
> bound or does it require high bandwidth?
> 2. Is it possible to get information about CPU usage or memory 
> consumption of a planet-lab node during some period in the past in order 
> to identity the behavior of the applications competing with my own for 
> the resources in the node planetlab2.cs.virginia.edu?
> 
> I am looking forward to hearing from you...
> Lívia
> 
> _______________________________________________
> Users mailing list: Users at lists.planet-lab.org
> https://lists.planet-lab.org/mailman/listinfo/users