[OpenAFS] fs setserverprefs and supernetting
Tue, 2 Nov 2010 09:57:54 +0000
The real issue (I've found) is to be able to distinguish between the=20
structural delay across a path and that part of the delay that is due
to the statistical multiplexing (contention for the common resources).
The first is going to be there come what may (barring a topology =
the second it what varies moment by moment. You can extract these values
by observation and some calculation (not too difficult).
Of course if the limiting link rate is different on different routes =
you may need to consider that effect as well (though that can also be
observed and calculated).
On 2 Nov 2010, at 08:10, Assarsson, Emil wrote:
>> -----Original Message-----
>> From: Jim Rowan [mailto:email@example.com]=20
>> Sent: m=E5ndag den 1 november 2010 16:54
>> To: Assarsson, Emil
>> Cc: firstname.lastname@example.org
>> Subject: Re: [OpenAFS] fs setserverprefs and supernetting
>> On Nov 1, 2010, at 4:31 AM, Assarsson, Emil wrote:
>>> Hi all,
>>> I would just like to have your input on this method of setting =20
>>> serverprefs based on the supernet distance.
>>> The background is that we have a large WAN with high latencies. The =20=
>>> default method of calculating the weight for serverprefs is not =
>>> We could have done this with static weighting but we need to keep =20=
>>> things really flexible and self-configuring.
>> There's a built-in assumption here that IP address assignments have a =
>> clear relationship with latency. That's probably accurate in so far =20=
>> as your company uses supernetting for routing. If you hand off Wan =20=
>> connections to MPLS, for instance, that might no longer be a good =20
>> assumption, and this approach may fall down. This algorithm does =20
>> appear to be an improvement over the built-in one, though.
> I guess that periodic latency measurements and comparisons between =
servers would be better but that is just too hard to write in perl :-P
> OpenAFS-info mailing list