FunctionCallBenchmark

I am comparing the latency of three types of function calls on a multi-core processor: 1) virtual, 2) normal, 3) inline (no call). The last one is the reference. In the benchmark I delibarately pin the threads to specific CPU-cores to avoid context switching since this would introduce unpredictable overhead and therefore make the results less reliable. I measure the virtual call latency when running on a single thread, or multiple threads, which are either pinned on the same physical core or not. The processor is heterogeneous as it contains two types of cores: 1) performance (two hardware threads) and 2) efficient cores (one hardware thread). In all experiemnts the virtual call latency was consitently between 0.8 and 1 ns. The normal function call did not show any measurable latency.

10-core 12th Gen Intel(R) Core(TM) i7-12650H - g++ 11.4.0 - ubuntu 22.04.2

logical core 0:

• Thread 0 timings: Virtual call: 6.48 ns, Normal call: 5.66 ns, Inline call: 5.64 ns

logical cores 0 and 1 (two threads - one physical core):

• Thread 0 timings: Virtual call: 9.86 ns, Normal call: 9.03 ns, Inline call: 9.04 ns

• Thread 1 timings: Virtual call: 9.85 ns, Normal call: 9.03 ns, Inline call: 9.04 ns

logical cores 0 and 2 (two threads - two physical cores):

• Thread 0 timings: Virtual call: 6.51 ns, Normal call: 5.67 ns, Inline call: 5.67 ns

• Thread 1 timings: Virtual call: 6.50 ns, Normal call: 5.67 ns, Inline call: 5.66 ns

Virtual function has a 12-13% overhead compared to a normal function call when running on a single thread. Multithreading on the same physical core makes each call approximately 30% slower and at the same time hides some of the virtual call latency - the overhead drops to 8%. Multithreading on two physical cores shows the same overhead as without multithreading. In absolute time the v-table lookup cost is in the order of 1 ns.

logical core 11 (performance-core) Vs logical core 12 (efficient-core):

• Thread 11 timings: Virtual call: 6.40 ns, Normal call: 5.64 ns, Inline call: 5.63 ns

• Thread 12 timings: Virtual call: 9.28 ns, Normal call: 8.42 ns, Inline call: 8.43 ns

The processor has 6 performance cores and 4 efficient cores. Each performance core has two logical cores (so-called hardware-threads, or, hyperthreads), these correspond to logical cores 0-11. Each efficient core has 1 logical core, these correspond to logical cores 12-15. The comparison between logical core 11 and 12 is a comparison between a performance and an efficient core. Performance cores are approximately 30% faster than efficient cores (turns-out "efficient" cores are not so efficient). The v-table overhead is 0.8 ns in both cases (12.5% on performance cores - 10% on efficient cores).