Performance of the Network Cache

next up previous
Next: Performance of the Up: Performance Results Previous: Performance of Communication

Performance of the Network Cache

The simplest measure of Network Cache performance is the hit rate, defined as,

and shown in Figure 15. (Note that local interventions are counted in the numerator.) Retries are generated locally when a cache line is locked in the NC due to a pending remote request. This locking could be due to a request to the same or a different cache line (cache conflict) from another processor. However, given the large size of the NC and the fact that there can only be 4 outstanding requests at one time (because each R4400 processor can generate only one request at a time), the chances of such a conflict are slim. Most retries are due to concurrent requests to the same cache line. When the pending request returns through the network and unlocks the line, the next retry will succeed. (Assuming the line is not ejected in the interim, which is unlikely.) This masking-out of simultaneous requests is termed the 'combining' effect, since multiple requests result in only a single network access. This effect is displayed in Figure 16.

Another reduction in network traffic is gained from what is termed the 'migration' effect. In essence, when data brought onto a station is then accessed by another processor, a remote access is potentially saved. This is true both for data that is dirty, as well as data that is shared. It is worthy of note that a system utilizing bus snooping would also see this benefit, but only for dirty data.

Figure 17: Average utilization of communication paths.

Figure 18: Local and central ring interface delays.

Stephen D. Brown
Wed Jun 28 18:34:27 EDT 1995