Costs of new primitives

We measured the costs of primitive operations on resource containers. For each new primitive, a user-level program invoked the system call 10,000 times, measured the total elapsed time, and divided to obtain a mean ``warm-cache'' cost. The results, in Table 1, show that all such operations have costs much smaller than that of a single HTTP transaction. This implies that the use of resource containers should add negligible overhead.

 

Table 1: Cost of resource container primitives.

Operation	Cost (us)
create resource container	2.36
destroy resource container	2.10
change thread's resource binding	1.04
obtain container resource usage	2.04
set/get container attributes	2.10
move container between processes	3.15
obtain handle for existing container	1.90

 

We verified this by measuring the throughput of our server running on the modified kernel. In this test, the Web server process created a new resource container for each HTTP request. The throughput of the system remained effectively unchanged.