Stefan Leigh, Andrew Rukhin
Kevin Mills, Virginie Galtier, Mudumbai Ranganathan
In current network switchers and routers, per-packet processing requirements in a network node are fairly homogeneous, and well-characterized system-independent metrics exist for bandwidth and memory resources. The advent of ``smart'' or so-called active networks will change that situation. Active networks will require an acceptable system-independent means of expressing CPU time requirements to enable the allocation and management of CPU cycles among active network nodes.
Any effective metric for CPU time usage in an active network node environment must account for at least 5 main sources of variability: (1) node hardware performance, (2) the specific execution environment in which the active application executes, (3) the mapping of the node OS system calls to real system calls in the host OS, (4) the implementation of the real system calls within the host OS, and (5) the behavior of the active application itself. An active network node model is proposed to account for the first four sources of variability. It makes use of simple proportions among local nodes and reference nodes with respect to benchmark workloads for each node OS call. A continuous-time finite-state (semi-)Markov chain model is proposed for the behavior of the active application itself.
Ultimately, the statistic(s) of interest in the active application model will be high quantile estimates for first return times to initial - idle - states. But a natural first step is to collect data on system call dwell (execution) and transition times to check the appopriateness of the proposed Markov model. To this end, a taxonomy of existing computer network performance benchmark software - synthetic, real, and hybrid - has been developed, examples are being collected, selected scenarios are being run, and their dwell/transition time profiles analyzed.
The Figure depicts the approximate exponentiality of the dwell-plus-transition time distributions for about half of the state-to-state transitions for one scenario from the ANTS execution environment PING application. The other half of the state-to-state transitions seem to be characterized by mixtures of exponentials with a wide range of dispersions.
This work is funded in part by DARPA.
Figure 2: Exponential probability plots for state-to-state transitions for an ANTS PING scenario. Except for a few outliers, the exponential distribution is a reasonable model for the cases shown.
Date created: 7/20/2001