fd_clock

By popular demand, here is an API for synchronizing various
cheap-inaccurate-weird-unit local tickcounters (e.g. fd_tickcount /
RDTSC, FPGA hardware timestamps, NIC hardware timestamps, etc) to
slow-accurate-normal-unit reference clocks (fd_log_wallclock /
clock_gettime / CLOCK_REALTIME, etc).

This allows users to accurately, cheaply, monotonically, concurrently,
consistently and deterministically convert various rando cheap tickcount
reads in HPC critical paths in a distributed application to an
equivalent of the system clock.

For example, using fd_clock on the pair fd_tickcount / fd_log_wallclock
for typical Linux x86 allows the fd_tickcount to act like
fd_log_wallclock with quantization limited sync accuracy (~<1 ns), a
couple of times lower latency (O(1) ns to convert RDTSC) with over an
order of magnitude small jitter (also O(1) ns to convert) and no
interthread sync error (shared memory interprocess persistent data
structure ... threads system wide can concurrently use a synchronization
lockfree parallel).

Because this requires FD_HAS_DOUBLE, this is considered an optional util
and must be included explicitly to use.

Note that fd_tickcount / fd_log_wallclock were given fd_clock_func_t
compatible variants.

IMPORTANT SAFETY TIP!  When compiling for non-x86 targets, fd_tickcount
by default falls back fd_log_wallclock.  While fd_clock will still
synchronize the pair fd_tickcount / fd_log_wallcount reasonably, it
really is just synchronizing the wallclock to itself with no practical
benefit.  Production non-x86 targets should provide an appropriate
target specific implementation for fd_tickcount to benefit from this API
in this case.

Author: kbowers-jump

src/util/clock/Local.mk

@@ -0,0 +1,6 @@
+ifdef FD_HAS_DOUBLE
+$(call add-hdrs,fd_clock.h)
+$(call add-objs,fd_clock,fd_util)
+$(call make-unit-test,test_clock,test_clock,fd_util)
+$(call run-unit-test,test_clock)
+endif