KDC CPU Loading

Throughput issues related to KDCs in PacketCable and CableHome networks are widely misunderstood. Here is a paper that describes what KDC throughput is actually necessary in deployed PacketCable networks.

Before any software release, we typically run our KDCs against one or more large networks of simulated MTAs (most of which are well-behaved, but some of which are attempting to crash the KDC). The graph below shows typical results from such a run.

The graph was created by setting up a network of 1,000,000 or more simulated MTAs, and then running a single KDC for a period of 24 hours against this network. Throughput on the KDC is CPU-limited (because of the public key operations that it must execute); issues such as memory and disk speed/storage are unimportant, so the graph is an accurate indicator of the total loading on the KDC is as a function of the size of the network. The individual data points were acquired by taking a brief snapshot of the KDC process (and all its child threads) once per minute.

The hardware used to generate the graph was an off-the-shelf inexpensive PC running Mandrake Linux 9.0. The PC has a single processor, a 1.8 GHz Athlon. Performance on other computer configurations increases roughly linearly with the number of processors and/or the CPU clock speed. We note that the measured CPU load in the graph scales as a function of the number of MTAs exactly as described in our theoretical paper on the subject.