Search squid archive

Re: Fastest peer without ICP ?

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 30/03/11 00:58, Yucong Sun (åéé) 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


[Index of Archives]     [Linux Audio Users]     [Samba]     [Big List of Linux Books]     [Linux USB]     [Yosemite News]

  Powered by Linux