3장. Improving the network latency

The consumption state (C-states) of CPUs optimize and reduce the power consumption of computers.

The C-states are numbered, starting at C0. In C0, the processor is fully powered and executing. In C1, the processor is fully powered but not executing. The higher the number of the C-state, the more components the CPU turns off.

Whenever a CPU core is idle, the built-in power saving logic steps in and attempts to move the core from the current C-state to a higher one by turning off various processor components. If the CPU core must process data, Red Hat Enterprise Linux (RHEL) sends an interrupt to the processor to wake up the core and set its C-state back to C0.

Moving out of deep C-states back to C0 takes time due to turning power back on to various components of the processor. On multi-core systems, it can also happen that many of the cores are simultaneously idle and, therefore, in deeper C-states. If RHEL tries to wake them up at the same time, the kernel can generate a large number of Inter-Processor Interrupts (IPIs) while all cores return from deep C-states. Due to locking that is required while processing interrupts, the system can then stall for some time while handling all the interrupts. This can result in large delays in the application response to events.

           Pkg(HW)  |            Core(HW) |            CPU(OS) 0   CPU(OS) 4
                    |                     | C0 active   2.5%        2.2%
                    |                     | POLL        0.0%    0.0 ms  0.0%    0.1 ms
                    |                     | C1          0.1%    0.2 ms  0.0%    0.1 ms
C2 (pc2)   63.7%    |                     |
C3 (pc3)    0.0%    | C3 (cc3)    0.1%    | C3          0.1%    0.1 ms  0.1%    0.1 ms
C6 (pc6)    0.0%    | C6 (cc6)    8.3%    | C6          5.2%    0.6 ms  6.0%    0.6 ms
C7 (pc7)    0.0%    | C7 (cc7)   76.6%    | C7s         0.0%    0.0 ms  0.0%    0.0 ms
C8 (pc8)    0.0%    |                     | C8          6.3%    0.9 ms  5.8%    0.8 ms
C9 (pc9)    0.0%    |                     | C9          0.4%    3.7 ms  2.2%    2.2 ms
C10 (pc10)  0.0%    |                     |
                    |                     | C10        80.8%    3.7 ms 79.4%    4.4 ms
                    |                     | C1E         0.1%    0.1 ms  0.1%    0.1 ms
...