atop

Kapil Thangavelu kapil.thangavelu at canonical.com
Thu Jul 15 21:09:55 UTC 2010


On Tue, 13 Jul 2010 14:30:02 -0400, Jos Boumans  
<jos.boumans at canonical.com> wrote:

>
> On 30 Jun 2010, at 17:45, John Tapsell wrote:
>>> Gimme an example of a problem that atop will help solve for which no
>>> other method will suffice.
>>
>> For network usage:
>> * The internet is slow but you don't know what program is
>> downloading/uploading.  You run atop and see that the ftp server is
>> uploading at 1MB/s.
>
> This is huge, and I don't know of any other tool that can do
> this and know several people that use the tool for exactly
> this to diagnose network related issues.
>
> The other is the resource utilization logging. Very useful
> for diagnosing after the fact. Less critical than the one
> above, but for this, I also don't know of another tool that
> provides this functionality.
>
> Do both of these features require the kernel patches to be
> applied?
>
> Cheers,
>
> -Jos
>

The per process network activity requires the patches. The disk activity  
per process falls back to a less accurate mechanism without the patches.  
The rest of atop works fine without the patches. Its perhaps useful to  
keep in mind that atop, gets all of its process info from parsing the  
/proc filesystem. Andrew morton proposed that the info that patch yields  
should instead go into the task accounting facilities also used by  
cgroups. To date though the network activitiy accounting hasn't been  
incorporated although it has been implemented  
(http://cgrouphacking.blogspot.com/). The deficiencies that Andrew notes  
on the patches wrt to accuracy of blkio counters are just as relevant in  
their task accounting implementation (which has been incorporated), afaics.

There's a nice article on lwn, on atop and its functionality as compared  
to other tools
http://lwn.net/Articles/387202/

Fwiw, i think its great functionality from a monitoring perspective, but  
the patches are 'crack' mostly in that their expedient tools for folks  
trying to solve and monitor real world performance issues. I think its  
unlikely that atop will use the task accounting implementation as it would  
effectively mean a rewrite of  its stats gathering to use netlink instead  
of proc file parsing, and one which still (two years later) does not  
implement the same functionality. And atop is also effectively a single  
maintainer with no upstream repo... which by itself is a decent reason not  
to incorporate its kernel patches, as effectively it lacks a community of  
developers supporting these patches, and that burden would likely need to  
be taken up by ubuntu.

cheers,

Kapil







More information about the ubuntu-server mailing list