On Thu, Mar 22, 2012 at 3:41 PM, Tushar Krishna <[email protected]>wrote:
> ** > Hi Dean, > My answers are inline. > > - Tushar > > > On 03/22/2012 05:31 PM, Dean Michael Ancajas wrote: > > Hi Tushar, > Thanks for the prompt reply. I have a follow up > question/clarifications though. Let's say I have successfully deflected a > HEAD_ flit to a different output port (from what was calculated in the RC > stage) all the BODY_ and TAIL_ flits will automatically follow right since > they would use the same input VC? At least that is how I understood the > code in InputUnit_D::wakeup(). > > As long as you make sure that the new output port is updated in the > "route" variable in that VC after the head deflects, the BODY and TAIL will > follow that same output port. > > > > Another question is, if we have deflected a flit to a non-optimal > path, is there a way to update its new path without diving into the details > of how the path is embedded into the message during the > NetworkInterface::flitisize ( ) part? What I would want to do is update the > shortest path to the source that is embedded in the flit/network message > based on its current deflected location. If I don't update the path/NetDest > inside the flit i think it would just go back to the router where it came > from since that would be the only node that would match in its routing > table. > > Hmm depends. Suppose a flit could go out of N or E for minimal routing > (and E had been computed by RC for XY routing). You deflected it out of S. > Now at the new router, it can either go out of N to come back to the old > router, or it can go out of E. I think it will go out of E if XY routing is > enforced (if you use a Mesh, all X links have a weight of 1 and Y links > have a weight of 2 to enforce XY routing). > But on the other hand, if a flit could only go out of N and you deflected > it to W, S or E, it will come back to the same router. > > This last case is what I am actually worried about since it would have a pingpong effect whichever way I deflect it. I'd also like to ask if there is an online guide on how to enable the debug mechanisms embedded in the garnet code? The instructions included in the ruby README file apparently does not actually work for garnet. > > > cheers! > dean > On Thu, Mar 22, 2012 at 12:40 PM, Tushar Krishna <[email protected]>wrote: > >> Hi Dean, >> The OutputUnit does not wake up when the flit has to leave the router (it >> only wakes up to receive credits). >> It is the switch that wakes up and schedules the output link. >> >> The SWallocator_d removes winning flits from their respective input VCs >> and places them in the switch buffer. >> The Switch_d reads from the switch buffer and pushes flits into the >> out_buffer of the particular output unit. >> The NetworkLink_d wakes up and reads flits out of the output buffers. >> >> If you want to implement locking, I think you will need changes in the >> SwitchAllocator_d where input flits place requests for output ports and >> output ports arbitrate among these requests. >> You need ensure that an output port only favors packets of the same input >> VC, before it can serve other VCs/ports. >> The Switch and Link are dumb units that just transfer flits. >> >> - Tushar >> >> >> >> On 03/22/2012 09:47 AM, Dean Michael Ancajas wrote: >> >> Hi, >> I want to implement deflection routing in garnet. At the >> InputUnit_d::wakeup() I lock the output port of the router (everytime a >> HEAD flit or HEAD_FLIT_ is encountered) so that no other flit will be >> scheduled thru this output port. My question is, when do I release the >> lock? Is it in the OutputUnit_d::wakeup() when a TAIL flit is about to >> traverse the link (i.e. when it puts the flit on the network link)? >> >> -dean >> >> >> _______________________________________________ >> gem5-users mailing >> [email protected]http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >> >> >> _______________________________________________ >> gem5-users mailing list >> [email protected] >> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >> > > > _______________________________________________ > gem5-users mailing > [email protected]http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users > > > _______________________________________________ > gem5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >
_______________________________________________ gem5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
