Max throughput of Random traffic

[zliu2014]

Hi,

I got the max throughput of 2d 6x6 Mesh with Random traffic is about 0.07 flits/node/cycle for 1 VC (4 buffers).
Is this an expected result ? Thanks

---Z. Liu

[zliu2014]

// This is the said test configuration: class MyNoC6x6Config extends NoCTesterConfig(NoCTesterParams(NoCParams( topology = Mesh2D(6, 6), channelParamGen = (a, b) => UserChannelParams(Seq.fill(1) { UserVirtualChannelParams(4) }), ingresses = (0 until 36).map { i => UserIngressParams(i) }, egresses = (0 until 36).map { i => UserEgressParams(i) }, flows = Seq.tabulate(36, 36) { (s, d) => if (s != d) Some(FlowParams(s, d, 0)) else None }.flatten.flatten, routingRelation = Mesh2DDimensionOrderedRouting() )))

[jerryz123]

How are you measuring throughput? When the offered traffic exceeds the capacity of the network, throughput will plummet. Here is an example result from PPIN:

[zliu2014]

Hi, Jerry. Thanks for the valuable info. I increased the ROB entry size to 1024 in TestHarness.scala, and make sure each Node (Ingress node) can only allocated up to 16 entries (16 * 36 < 1024), and the packet size is 1 flit. Then I run the test with random traffic for 10^6 flits totally. I print the throughput using "total_flits/total_cycles/36", as "flits per cycle per node". Based your above graph, my test probably corresponds to the decreasing region ? What is the definition of the "Offered traffic (fraction of capacity)" in the above figure ?

[jerryz123]

From PPIN:

The capacity is defined as the best possible throughput, assuming perfect routing and flow control, that could be achieved by the network under the given traffic pattern.

You may want to look at running the EvalHarness classes, those will throttle traffic and measure the throughput for you.