pcp
[Top] [All Lists]

Re: [pcp] RFE: pmcd client tracking

To: Nathan Scott <nscott@xxxxxxxxxx>
Subject: Re: [pcp] RFE: pmcd client tracking
From: Max Matveev <makc@xxxxxxxxx>
Date: Sat, 18 Apr 2009 11:10:28 +1000
Cc: pcp@xxxxxxxxxxx
In-reply-to: <556086237.3539271239952360205.JavaMail.root@xxxxxxxxxxxxxxxxxx>
References: <1726636914.3539251239952341295.JavaMail.root@xxxxxxxxxxxxxxxxxx> <556086237.3539271239952360205.JavaMail.root@xxxxxxxxxxxxxxxxxx>
>>>>> "nscott" == Nathan Scott <nscott@xxxxxxxxxx> writes:

 nscott> I was asked today if we could add an option to the pcp(1) script to
 nscott> report which clients are connected to a host.  We'd need pmcd.client
 nscott> metrics to do that ... pmcd has a table with connected clients, but
 nscott> its not exported currently.  It also only has IP addresses, it would
 nscott> be better if we had hostnames, and better still if we had pmProgname
 nscott> sent over from the client, and perhaps a PID too (would need to add
 nscott> to the CREDS PDU to pass those over).

PIDs would be relatively simple to add - you may need to split them
across 2 CREDs to fit all 32 bits but otherwise it's just binary.

progname is more difficult - you'd have to force all clients which
link with libpcp to provide it with one and you'd need to find a way
to pass it across which would almost certainly mean a protocol
revision or some creating shuffling of bits across multiple CREDs.
Plus is could be seen as "invasion of privacy". In short, I wouldn't
do it if I was in your shoes.

And for hostnames, be very affraid of the name lookups - bind has 45
second timeouts which means that any getfoobybar can take up-to this
much time, with pmcd's single threaded nature it could mean that it
will be out of action a lot. If you want to have hostnames, it's
better to implement the lookups client side.

max
 

<Prev in Thread] Current Thread [Next in Thread>