miller/docs/src/cpu.md.in
2022-01-10 19:55:01 -05:00

57 lines
2.7 KiB
Markdown

# CPU/multicore usage
Miller 6 is written in [Go](https://golang.org/) which supports multicore programming.
Miller uses Go's _channel_ concept. The following are all separate,
concurrently executing "goroutines", which loosely speaking you can think of as
"threads":
* Two for the input-record reader: one to read files/pipes into memory as byte streams; another parsing those byte streams to record objects
* One for each [verb](reference-verbs.md) in [then-chains](reference-main-then-chaining.md)
* One for output-record writer: formatting output records as text
* One controller which coordinates all these, without much work to do.
For example, `mlr --csv cut -f somefield then sort -f otherfield then put '$z =
$x + $y' a.csv b.csv c.csv` will have seven goroutines running:
* record-reader byte-reader,
* record-reader record-builder,
* `cut`,
* `sort`,
* `put`,
* record-writer,
* controller.
Ideally all seven of those would utilize 100% of a CPU, for 700% total. This
is not the general case, for a few reasons:
* If all the verbs in the chain are [streaming](streaming-and-memory.md) --
operating on each record as it arrives, then passing it on -- then all verbs
in the chain will be active at once.
* On the other hand, if there is a [non-streaming verb](streaming-and-memory.md)
in the chain, which produces output only after receiving all input -- for
example, `sort` -- then we would expect verbs after that in the chain to sit
idle until the end of the input stream is reached, the `sort` does its
computation, then sends its output to downstream verbs.
* In practice, profiling has shown that the record-readers and record-writers
use the most CPU of all the above. This means CPUs running verbs may not be
100% utilized, since they are likely to be spending some of their time waiting
for input data.
Note that goroutines aren't pinned to CPUs; the Go runtime schedules them, and
the operating system schedules multiple executables running at the same time.
In practice, total CPU usage runs to at most 250% or maybe 350% for deep
processing chains, and the system is free to allocate CPU where it is needed.
Since the number of CPUs Miller will use is tied to the number of verbs in the
processing chain, running Miller on a machine with more CPUs than active
routines (as listed above) won't speed up a given invocation of Miller.
However, of course, you'll be able to run more invocations of Miller at the
same time if you like.
You can set the Go-standard environment variable `GOMAXPROCS` if you like. If
you don't, Miller will (as is standard for Go programs in Go 1.16 and above) up
to all available CPUs.
If you set `$GOMAXPROCS=1` in the environment, that's fine -- the Go runtime
will multiplex different channel-handling goroutines onto the same CPU.