[Click] dynamic ticket adjustments in FromDevice and ToDevice

[Eddie Kohler]

Hi Frederic,

I think you have now graduated :) but I wanted to respond to this mail.

Your analysis of CLICK_DEVICE_ADJUST_TICKETS isn't quite on the money, since 
that symbol is always defined.  But -- well, here's a comment from FromDevice.

     // 9/18/06: Frederic Van Quickenborne reports (1/24/05) that ticket
     // adjustments in FromDevice+ToDevice cause odd behavior.  The ticket
     // adjustments actually don't feel necessary to me in From/ToDevice any
     // more, since FromDevice's interrupt handler will reschedule FromDevice
     // as necessary; now "ticket adjustment" is subsumed by "scheduled or not
     // scheduled".  So commenting this out.

At this point FromDevice does not adjust its tickets, and non-polling-mode 
ToDevice does not adjust its tickets either.

If anyone sees weird behavior as a result, please LET THE LIST KNOW!

Thanks much,
Eddie


Frederic Van Quickenborne wrote:
> Hi all,
> 
> for some tests we used the click modular router as a gateway between a 
> 100 Mbit/s and a 10 Mbit/s network. But we had lot's of troubles 
> understanding the resulting packet loss and delays of the packets, e.g., 
> the first test after installing the click configuration differed from 
> later tests. Then we found a possible problem as the tickets for the 
> fromdevice were as high as 1024 and the tickets for the todevice were 
> very low. Then we used the nostride option when configuring click, but 
> nothing changed.
> 
> We compared the code around the adjust_tickets call in the fromdevice 
> and the todevice, and we found this:
> 
> in fromdevice.cc, line 324:
> <code>
> #if CLICK_DEVICE_ADJUST_TICKETS
>     adjust_tickets(npq);
> #endif
> </code>
> 
> in todevice.cc, line 286:
> <code>
>     adjust_tickets(npq);
> </code>
> without any #if CLICK_DEVICE_...
> 
> Is it possible that the #if and the #endif lines were forgotten in the 
> todevice.cc file?
> 
> Anyway, after deletion of the adjust_tickets lines, the packet loss and 
> the delays were quite predictable (but Click is a lot more performant 
> when using the automatic tickets adjustment and when using polldevices).
> 
> Kind regards,
> Frederic.
> 
> 
> ==================================================================
> Frederic Van Quickenborne
> Department of Information Technology (INTEC)
> Ghent University - IMEC
> Sint-Pietersnieuwstraat 41, B-9000 Gent, Belgium
> tel.: +32 (0)9-264 9957; tel. secr.: +32 (0)9-264 9970
> fax: +32 (0)9-264 9960
> Frederic.VanQuickenborne at intec.ugent.be
> http://www.ibcn.intec.ugent.be
> 
> _______________________________________________
> click mailing list
> click at amsterdam.lcs.mit.edu
> https://amsterdam.lcs.mit.edu/mailman/listinfo/click