Hi, Peter, If you have enough ports on the switch then you certainly can configure things to send the packets directly from the ROACHs to the various X boxes. The pre-built PAPER model will support this if properly configured. The one thing to keep in mind is the IP-to-MAC address table in the 10 GbE cores. These are setup by paper_feng_init.rb.
You could use the "factory" MAC addresses of the X-box 10 GbE interfaces or you could configure the X boxes to set their 10 GbE MAC addresses to predefined values. If using the latter approach, we often use 02:02:ww:xx:yy:zz where ww:xx:yy:zz corresponds to the IP address. Hope this helps, Dave On Dec 1, 2014, at 5:21 AM, Peter Niu wrote: > Hi,Dave, > Thanks for your Document about EQ,and suggestion about the sample rate.Now I > have a question about the correlator architecture. > I have saw your PPT : > Correlator Architectures > Present and Future > CASPER Workshop 2011 > > The structure mentioned in the PPT is the structure PAPER used now .Using a > set of precise IP assignment to avoid Loop Back is ok. However,If we use the > Packetized F/X Concept:Uses two ports on switch per F/X pair.It may not meet > the Loop Back problem.On the switch, The IP address will tell the packet > which Xeng to go .The structure which The PAPER model using now is the > eth_?_gpu port on ROACH connect HPC port directly.Is this only for saving > ports on switch?Well,Our switch have 64 ports,If we use the two ports on > switch per F/X pair Concept,the ports may be sufficient . > This is the question asked by my teacher Wu fengquan. As PAPER provide a lot > of ruby control scripts online to use,I'd rather use this model exits > now.What should I say to him?Is there some more advantages to use this > structure instead the two ports on switch per F/X pair? > Thanks for your help! > Best wishes! > Peter > > > >
