Garnet NoC: Flit Divertion
Hi everyone,
I'm simulating a unique architecture in Garnet NoC, where I designed functions in the Input Unit, Router, and Network Interface to divert flits from their original source and destination to specific ones (by tampering with the source and dest ID of flits). I expect the routing path of modified flits to change due to altered routing information. However, when I rebuild and run the traffic pattern benchmark on this new architecture, the buffer utilization, network latency, link utilization, and flits/packets read/write of each router remain the same as the results from the normal architecture (default Garnet NoC), though the host running time slightly differs. Can anyone explain why the flit diversion isn't successful? I'd like to see changes in the performance statistic data in the figure attached. Especially for buffer reads/writes, I expect some to increase and others to decrease due to the different routing paths. However, they are currently identical.
Many thanks!
[cid:e71faa8b-ec3a-4530-a228-4eff14c35afd]
Hi Haoyu,
Are you changing the RouteInfo of the head flit? If so, it should reflect
the change. But make sure you really touch the RouteInfo object of flit but
not a temporally created object from the get_route().
Best,
Jiayi
On Tue, May 9, 2023 at 8:32 PM Haoyu Wang via gem5-users <
gem5-users@gem5.org> wrote:
Hi everyone,
I'm simulating a unique architecture in Garnet NoC, where I designed
functions in the Input Unit, Router, and Network Interface to divert flits
from their original source and destination to specific ones (by tampering
with the source and dest ID of flits). I expect the routing path of
modified flits to change due to altered routing information. However, when
I rebuild and run the traffic pattern benchmark on this new architecture,
the buffer utilization, network latency, link utilization, and
flits/packets read/write of each router remain the same as the results from
the normal architecture (default Garnet NoC), though the host running time
slightly differs. Can anyone explain why the flit diversion isn't
successful? I'd like to see changes in the performance statistic data in
the figure attached. Especially for buffer reads/writes, I expect some to
increase and others to decrease due to the different routing paths.
However, they are currently identical.
Many thanks!
gem5-users mailing list -- gem5-users@gem5.org
To unsubscribe send an email to gem5-users-leave@gem5.org
Hi Jiayi,
Yes, the RouteInfo modification was successful. Thanks for your reminder.
My issue has been resolved. The default routing algorithm searches the table for a fixed routing path, so the actual routing path isn't diverted by the new RouteInfo. When I switched to the XY routing algorithm, flits were diverted as expected.
Best regards,
Haoyu
From: Jiayi Huang jyhuang91@gmail.com
Sent: 09 May 2023 15:25
To: The gem5 Users mailing list gem5-users@gem5.org
Cc: Haoyu Wang Haoyu.Wang@soton.ac.uk
Subject: Re: [gem5-users] Garnet NoC: Flit Divertion
CAUTION: This e-mail originated outside the University of Southampton.
Hi Haoyu,
Are you changing the RouteInfo of the head flit? If so, it should reflect the change. But make sure you really touch the RouteInfo object of flit but not a temporally created object from the get_route().
Best,
Jiayi
On Tue, May 9, 2023 at 8:32 PM Haoyu Wang via gem5-users <gem5-users@gem5.orgmailto:gem5-users@gem5.org> wrote:
Hi everyone,
I'm simulating a unique architecture in Garnet NoC, where I designed functions in the Input Unit, Router, and Network Interface to divert flits from their original source and destination to specific ones (by tampering with the source and dest ID of flits). I expect the routing path of modified flits to change due to altered routing information. However, when I rebuild and run the traffic pattern benchmark on this new architecture, the buffer utilization, network latency, link utilization, and flits/packets read/write of each router remain the same as the results from the normal architecture (default Garnet NoC), though the host running time slightly differs. Can anyone explain why the flit diversion isn't successful? I'd like to see changes in the performance statistic data in the figure attached. Especially for buffer reads/writes, I expect some to increase and others to decrease due to the different routing paths. However, they are currently identical.
Many thanks!
[cid:18800e19deccb971f161]
gem5-users mailing list -- gem5-users@gem5.orgmailto:gem5-users@gem5.org
To unsubscribe send an email to gem5-users-leave@gem5.orgmailto:gem5-users-leave@gem5.org
