>AFAIK, in todays cache meshes, a cache must know its peers. It may
be
>reasonable to suggest that when asymmetric links and other
complications
>exist, a cache administrator should configure peer selection
algorithm
>appropriately. IMO, the described problem has little to do with
ICP. It
>is a question of how configurable peer selection is.
To solve this problem, a peer selection algorithm should know at
least the size of the requested URL which should be sent back in the
ICP_OP_HIT message. However, RFC 2186 gives no indication of the object size
in the ICP_OP_HIT message. So any peer selection algorithm won't work well
without this information that ICP should return.
>I do not know if ICP contains peer selection part (e.g., go to the
first
>HIT peer). Even if it does, the practice shows that peer selection
is
>separate from object discovery protocols (ICP, HTCP, Carp, Cache
Digest,
>etc.). Peer selection procedure drives object discovery routine(s),
>determining the order in which the latter are used, the time to
stop,
>and the final selection among several candidates.
Agreed ! The issue may concern other object discovery protocols. I
guess it is a good practice to separate object discovery protocols from peer
selection; however, object discovery protocols should at least give enough
information (such as object size) to the peer selection algorithm in order
to make the right choice.
>Real life networks are probably too complex to design a universal
>automated peer selection algorithm, no matter how much info that
>algorithm may have access to. This is especially true when people
are
>killing themselves over 1% increase in hit ratio.
One of the main goals of the caches is to reduce latency to improve
the QoS from the user point of view; IMO, some work has to be done (within
the WG) to determine what should be the right level of information to
transmit between the layer of object discovery protocols and the layer of
peer selection in order to fully reach this goal.
Ivan,
This archive was generated by hypermail 2b29 : Thu Nov 18 2004 - 11:21:26 MST