Re: [squid-users] Fastest peer without ICP ?

From: Amos Jeffries <squid3_at_treenet.co.nz>
Date: Wed, 30 Mar 2011 01:59:16 +1300

On 30/03/11 00:58, Yucong Sun (叶雨飞) wrote:
> it's a customized tunnel can only transport TCP data.
>
> on that diagram, SQUID BCD is actually act as a service cluster, for
> many client like A. the data between A and BCD must be encrypt, and
> the latency is expected to be pretty high on the link. flooding the
> limited bandwidth tunnel with HTCPs probably not a good idea either.
>
> However, per you, SQUID will track HTTP connect time and prefer faster
> one by default? can it track "response time" instead? which means
> time of a header exchange is done (which pretty much equals to the
> link speed).

The metrics there are very round-trip-time oriented at the packet level.
I know it measures the TCP handshake cycle. I have not spent much time
looking at those options to see if it adds later checkpoints as well.
Given the variation of header size from 15 bytes to 65 KB on any request
I very much doubt it is easily usable.

>
> But I don't understand how SQUID reset the peer preference though, in
> this case, if one of the peer is slow, squid wills till periodically
> query it just to make sure it is still slow, but that will be very bad
> experience for the end-user though.

Only the UDP and ICMP methods are actively probing.
The HTTP method of detection requires a request needing to use the link.
It is highly susceptible to periods of low or no requests and if a peer
gets detected completely dead there are problems detecting when it comes
live again (due to squid not sending HTTP through 'dead' peers to get
the measure). It is reliable measure only if the link is also reliably
available and traffic levels over it relatively often.

The preference will depends on the algorithm chosen to select peers. For
example weighted-round-robin will balance the load relative to the
apparent distance to peers (as measured in hops and/or time).

Whereas "closest" will prefer the closest one (in hops and/or time
again) until it starts to overload. Then the next closest will be used
as well. If the second is actually faster they switch places in a sudden
load swap. Things can get messy after that (from a traffic control
perspective, clients should not notice).

Amos

-- 
Please be using
   Current Stable Squid 2.7.STABLE9 or 3.1.11
   Beta testers wanted for 3.2.0.5
Received on Tue Mar 29 2011 - 12:59:21 MDT

This archive was generated by hypermail 2.2.0 : Tue Mar 29 2011 - 12:00:02 MDT