Context switches are the main cause of both latency and CPU load. We can distinguish three types of context switches: (1) rescheduling, requiring context switches between processes or between threads, (2) kernel calls, requiring context switching between process and kernel space, and (3) interrupts requiring context switching between process and kernel space.
In a general purpose OS, these context switches require a lot of CPU
time, and thereby cause both latency and load. Using light weight
alternatives for the OS context switches can therefore result in a
significant performance improvement. However, a light weight
alternative usually implies a restriction in some other field, e.g.,
no protection between processes, a limitation in the use of OS
functionality, or support for only a single process. If the
restriction is acceptable, this method offers an interesting
solution. Solutions that apply this method can be found in
[4], [14], and [20]. Graph C in
Figure 12 shows the potential of such an
improvement, assuming the 0-byte latency 
is 20 
s,
instead of 
, and assuming the interrupt handling
overhead can be completely removed, which reduces the interrupt
handling time from 32.5 s to 11.5 s.
Furthermore, graph D in Figure 12 shows the
effect of applying both the improvement methods: off-loading and
low-latency context switches.