After reading the previous paper I had raised a concern that a centralized scheduler can cause problems and the authors in this paper discuss that removing the central scheduler can make the router more scalable. It is proposed that optics with zero power consumption can be used to place the fabric of a 100Tb/s router in a single rack without losing throughput guarantees.
The load balancing architecture provides an extra load-balancing stage that helps in spreading non-uniform traffic and making it uniform. Uniform arrival of packets have guaranteed throughput of 100% with fixed equal-rate switch that has virtual queues. The load balancing switch does not have a central scheduler but it requires a switch fabric of NxN that can be reconfigured for each packet transfer. To get rid of reconfigurations the switch fabric can be replaced with a fixed mesh of optical channels. Each fixed equal-rate switch can be replaced with NxN fixed channels. The mesh needs to be uniform if the traffic matrix is not known. Wavelength division multiplexing (WDM) can be used to reduce the number of fibers by having each linecard multiplex N WDM channels onto one fiber. FOFF (Full Ordered Frames First) is used to have packets leave the switch in order. Failure of linecards raises the need to scatter traffic in uniformly. The solution is to increase the number of paths M between local switches.
The architecture proposed in the paper removes the centralized scheduler and have a switch fabric that does not need to be reconfigured.
The paper was interesting to read but for me it was difficult to follow all the details given in the paper. I feel that I have not learned much out of this paper and there are lots of things that I have missed. Probably after tomorrow's lecture this paper will be more clear to me.
The load balancing architecture provides an extra load-balancing stage that helps in spreading non-uniform traffic and making it uniform. Uniform arrival of packets have guaranteed throughput of 100% with fixed equal-rate switch that has virtual queues. The load balancing switch does not have a central scheduler but it requires a switch fabric of NxN that can be reconfigured for each packet transfer. To get rid of reconfigurations the switch fabric can be replaced with a fixed mesh of optical channels. Each fixed equal-rate switch can be replaced with NxN fixed channels. The mesh needs to be uniform if the traffic matrix is not known. Wavelength division multiplexing (WDM) can be used to reduce the number of fibers by having each linecard multiplex N WDM channels onto one fiber. FOFF (Full Ordered Frames First) is used to have packets leave the switch in order. Failure of linecards raises the need to scatter traffic in uniformly. The solution is to increase the number of paths M between local switches.
The architecture proposed in the paper removes the centralized scheduler and have a switch fabric that does not need to be reconfigured.
The paper was interesting to read but for me it was difficult to follow all the details given in the paper. I feel that I have not learned much out of this paper and there are lots of things that I have missed. Probably after tomorrow's lecture this paper will be more clear to me.
No comments:
Post a Comment