mirror of
https://github.com/nosqlbench/nosqlbench.git
synced 2024-11-22 08:46:56 -06:00
82 lines
1.6 KiB
Plaintext
82 lines
1.6 KiB
Plaintext
@startuml
|
|
|
|
Participant "Calling\nThread" as t
|
|
Participant "Limiter\nlogic" as l
|
|
Participant "Allocated\nnanos" as a
|
|
Participant "Elapsed\nnanos" as e
|
|
Participant "Clock\nSource" as c
|
|
|
|
t -> l : acquire(nanos)
|
|
|
|
group allocate start time
|
|
l -> a : getAndIncrement(nanos)
|
|
activate a #black
|
|
note over l,a
|
|
**allocated** is an atomic accumulator
|
|
which represents scheduled time. Each
|
|
op causes it to be atomically incremented
|
|
by a time slice of nanos.
|
|
end note
|
|
a -> l : <scheduled_at>
|
|
deactivate a
|
|
end
|
|
|
|
group calculate delay (cached)
|
|
l -> e : get()
|
|
activate e
|
|
note over e
|
|
**elapsed** is an
|
|
atomic register
|
|
which caches
|
|
system time.
|
|
end note
|
|
e -> l : <elapsed>
|
|
deactivate e
|
|
l -> l : delay = \nelapsed - scheduled_at
|
|
|
|
note right
|
|
**delay** measures external delay
|
|
that causes an op to fire after
|
|
the ideal time. **positive delay**
|
|
thus means the rate limiter doesn't
|
|
need to impose its own blocking delay
|
|
in order to ensure delay>=0.
|
|
end note
|
|
|
|
end
|
|
|
|
group if delay<0 (cached)
|
|
note over l,c
|
|
If delay<0, then this operation is too soon according
|
|
to the cached clock value. Since this could be stale
|
|
and cause us to block needlessly, we update the cached
|
|
clock value and recompute delay.
|
|
end note
|
|
l -> c : get() (~25ns)
|
|
activate c #orange
|
|
c -> l : <elapsed>
|
|
deactivate c
|
|
|
|
l -> e : store(<elapsed>)
|
|
activate e #black
|
|
e -> l
|
|
deactivate e
|
|
l -> l : delay = \nelapsed - scheduled_at
|
|
|
|
group if delay<0 (updated)
|
|
l->l: sleep(-delay);\ndelay=0
|
|
note right
|
|
If delay is negative, we sleep
|
|
in the calling thread and
|
|
set delay=0
|
|
end note
|
|
activate l
|
|
deactivate l
|
|
end
|
|
|
|
end
|
|
|
|
l->t: <delay>
|
|
|
|
|
|
@enduml |