Simple click config is VERY slow on our system; also the PollDevice doesn't work on our tulip.

[Eric Freudenthal]

Thanks for the speedy response.  In sequence:

1. haven't tried user level; will do on monday
2. We see the same latency/variance with DelayUnqueue(0s).  We'll try it
with LONG delays on mon.
3. good idea.  Will try on mon.
4. same behavior if we don't pass the packets to linux (using
ArpQuerier/Responder instead; I tried this today).
(and)
5. pls post a tar of a compatibly configured /usr/src/linux.  Perhaps my
config irritates click?

Thanks,
Eric

Eric Freudenthal  // Courant Institute // New York University
719 Broadway, Room 712 // New York, NY // 10003
office: 212-998-3345 // cell: 917-279-6208

-----Original Message-----
From: Eddie Kohler [mailto:kohler at icir.org]
Sent: Friday, March 01, 2002 8:37 PM
To: Eric Freudenthal
Cc: click at amsterdam.lcs.mit.edu; Taozhao at Cs. Nyu. Edu; Vijay Karamcheti
Subject: Re: Simple click config is VERY slow on our system; also the
PollDevice doesn't work on our tulip.

> However, a that inserts 10 latency/bandwidth limiter blocks (see end of
this
> note) adds an amazing 10ms, with a very large variance (we often see
> 30-70ms).  As before, there is no network traffic other than the pings.

Some thoughts.

(1) What is the behavior of your latency/bandwidth limiter blocks at user
level? Can you device a user-level experiment that shows similar bad
behavior? It wouldn't shock me if one or more of DelayUnqueue and
BandwidthShaper were buggy.

(2) What if you had just one DelayUnqueue with 10 times the delay? What
happens there? (Why do you cascade the blocks anyway?)

(3) What happens if you have one DelayUnqueue with the normal delay, but
you also have 10 DelayUnqueue blocks just sitting there idling? Like this:

    Idle
        -> Queue(256)
        -> BandwidthShaper(100000000)
        -> SetTimestamp
        -> DelayUnqueue(0.00001s)
        -> Idle;

Does this still have high variance? Each of the DelayUnqueue elements (even
the idling ones) sticks itself on Click's task list. This experiment is
trying to determine whether the variance is due to task list issues.

(4) What happens if you don't send the packets to Linux? We have
experienced problems recently with the ToLinux element. Its locking
discipline is broken. We are working on fixing it.

Eddie