On 6/17/19 11:50 PM, Andrew Hanushevsky wrote:
> Hello Adrian & Oliver,
>
>> just to chime in - I have fallen into the very same trap :-(.
>> But that easily explains why I did not notice any load balancing effects.
> Yes, the issue here is that the "sched" directive is used by the
> redirector and the "perf" directive is used by the data server. That
> means, depending on how you structured your config file, catching the
> discrepancy is not assured. That said, yes, the documentation should
> make it clear that cms.perf needs to be specified on the data server
> side. if you want to load balance.
>
>> While I fully agree a working default would be good,
> The first issue is to where to put the script so that it could be
> uniformly invoked. Since there was no good place, we decided to not to
> make a default (that's explicitly stated in the documentation).
well, the current place is as good as any .. i do not see why should be
a special place.. and also it makes sense
/usr/share/xrootd/utils/
>> the script shipped with XRootD right now is mostly useless on modern
>> systems - it only works with CentOS 6 and older since it hardcodes the
>> expected >columns from netstat output,
> That's the second reason for there being no default. Load collection
> scripts are, by necessity, sensitive to the OS and OS version being
> used. Since we couldn't guarantee that the script would do anything
> useful for future OS's, making it the default didn't seem like a good
> idea. Frankly, I still think a default is not a good idea as you really
> need to verify the script works for your particular OS. I suppose we
> could make that more clear in the docs.
well, this decision created the fact that a advertised functionality
_silently_ is not working! which IMHO is not ok ..
i see only two options :
either use no matter how broken script as default to have a minimal
functionality AND warn more or less loudly that cms.perf was not
explicitly set with a script
or complain really loud in logs in both redirector and servers that the
the cms.sched is not properly setup (on servers complain about the lack
of cms.perf)
>> I'd be willing to fix this ot just use the kernel interface from
>> procfs: /proc/net/dev
> That would be greatly appreciated :-)
well, i had a head-start :)
https://github.com/adriansev/bin-scripts/blob/master/cms_monPerf
but i really do no like the vmstat usage, i will change to use something
directly like /proc/stat
also i do not know what is supposed to be the PGIO
Any comments/ideas/feedback about it?
Thanks!
Adrian
########################################################################
Use REPLY-ALL to reply to list
To unsubscribe from the XROOTD-L list, click the following link:
https://listserv.slac.stanford.edu/cgi-bin/wa?SUBED1=XROOTD-L&A=1
|