A Modernize Version of the Antique Telephone Architecture Christopher Francis & Bertha Wilhelm

Goal of the Paper 1.Scalable 2.Easily manageable 3.Fault-tolerant 4.Efficient data center network fabric A "plug-and-play" large-scale, data center network The old telephone architecture !!! 1.Existing routing and forwarding trees do not work with fat – tree 2.Scalability issues, If not properly programmed as the number of nodes increases the bandwidth o f the network decreases significantly (Ref)

Fat-Tree Topology Major drawback: Assumes one basic topology and it's not clear if this is the best topology or if it will be best for future data centers. These results show that fat-tree is not non-blocking.

Fabric Manager  They never discuss the bottle-neck in the fabric manager.  Solutions if fabric manager fails?  FM is a central component of PortLand. Even though it maintains soft state, its failure may result in many ARP requests in a short period of time which in turn can potentially cripple a network with that many hosts sharing the same broadcast domain.  The paper suggests that traditional ARP can be used as a fallback mechanism in case the FM fails. However, cache pollution or simply malfunctioning of the address resolution in the FM bringing the network down.  There is also an implicit assumption of trust on the end hosts  Malicious virtual machines can query for non-existing addresses to generate broadcast traffic.

The Evaluation  Limited to a small test bed, understandable, but some of the results obtained may change significantly in a large test bed. (Sec 5, Scalability)  “ We use measurements from our existing system to project the requirements of larger systems.”  The switches used are actually PCs with NetFPGA cards, it would have been nice to see a testbed with off-the-shelf switches, as the design goal states, to confirm that it is actually feasible.  The paper does not account for what occurs if there is a failure during migration of virtual machine.

The Evaluation  The paper does not show potential benefits for energy efficiency, power consumption. An important factor that new data center designs should incorporate.  The paper does not provide quantative analysis comparing the other architectures to PortLand.  TRILL and SEATLE both use flat addressing and general topology while Portland uses hierarchical addressing and fat tree.

UNACHIEVABLE “This is beyond the capacity of a single modern machine, but this also represents a relatively significant number of ARP misses/second.” Solution: Move the FM to small-scale cluster = decentralize FM.

Cost Analysis  Fat-Tree implementation?  Rewiring from existing architecture to fat-tree  Separate Control Network is needed for the communication between the FM and local switch  “modest” cost (Sec 4.2 )

Thank you !

Unsuccessful Goals of the Paper To create a 1.Scalable  2.Easily manageable  3.Fault-tolerant  4.Efficient data center network fabric 