ParaSeq - Parallel execution of Iterables
use ParaSeq;
# The 1-millionth prime number: 15485863
say (^Inf).&hyperize.grep(*.is-prime)[999999];
say (^Inf).&hyperize.grep(*.is-prime).skip(999999).head;
# Fetching lines of files, each element containing a List of lines
my @lines = @filenames.&hyperize(1, :!auto).map(*.IO.lines.List);
my %lines = @filenames.&racify(1, :!auto).map: { $_ => .IO.lines.List }
ParaSeq provides the functional equivalent of hyper
and race
, but re-implemented from scratch with all of the experience from the initial implementation of hyper
and race
in 2014, and using features that have since been added to the Raku Programming Language.
As such it exports two subroutines hyperize
and racify
, to make them plug-in compatible with the hyperize
distribution.
This module is now considered to be in BETA, as all features mentioned in the IMPROVEMENTS section have been implemented.
# The 1-millionth prime number using subroutine syntax: 15485863
say hyperize(^Inf).grep(*.is-prime)[999999];
# Start with 2000 element batches, and use max 10 workers
say (^Inf).&hyperize(2000, 10).grep(*.is-prime)[999999];
# Always work with a batch-size of 1
my @lines = @filenames.&hyperize(1, :!auto).map(*.lines.List);
Change the given Iterable
into a ParaSeq
object if the following conditions have been met:
- degree is more than 1
If the number of simultaneous workers is set to 1, there is no real parallelizetion possible. In which case hyperize
will simply return its first argument (the invocant when used with the .&hyperize
syntax).
- source is not exhausted with first batch
If the first batch already exhausts the source iterator, then there is little point in parallelizing. A Seq
that will reproduce the original source sequence will then be returned.
- source
The first positional argument indicates the Iterable
source of values to be processed. If the .&hyperize
syntax is used, this is the invocant.
- initial batch size
The second positional argument indicates the initial batch size to be used. It defaults to 16
. Will also assume the default if an undefined value is specified.
- maximum number of simultaneous workers
The third positional argument indicates the maximum number of worker threads that can be activated to process the values of the source. It defaults to the number of available CPUs minus 1. Will also assume the defailt if an undefined value is specified.
- :auto / :!auto
Flag. Defaults to True
. If specified with a False
value, then the batch size will not be altered from the size (implicitely) specified with the second positional argument.
- :catch / :!catch
Flag. Defaults to True
. If specified with a False
value, then any exception that is thrown by any of the parallel worker threads, will be re-thrown immediately, rather than being presented on STDERR at the end of the process.
my %lines = @filenames.&racify(1, :!auto).map: { $_ => .IO.lines.List }
There is very little functional difference between hyperize
and racify
. The only difference is that while hyperize
will guarantee that the order of the produced values is the same as their source values, racify
does not offer such a guarantee.
From a performance point of view, racify
may be a few percent faster than hyperize
, but one should generally not use it for that reason.
One situation to use racify
over hyperize
is when any results are expected as soon as they become available, without waiting for any results from a previous batch being delivered.
One should also note that the use of the last
statement in a grep
or map
may produce random results, as delivery of produced values is halted after the batch that was produced while the last
statement was executed, has been delivered.
These methods return a single value that could be considered a reduction of the hypered iteration. In alphabetical order:
Because the .are
method is already highly optimized, adding specific hypering would only slow down execution. So no specific hypering support has been added.
Because the .categorize
method produces a Hash, putting together the Hash
results of each batch in a correct manner, would be very complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this method at this point.
Because the .classify
method produces a Hash, putting together the Hash
results of each batch in a correct manner, would be very complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this method at this point.
An optimized version of the .elems
method has been implemented.
An optimized version of the .end
method has been implemented.
The .first
method is supported without any specific hypering support added as it turned out to be impossible to make it faster.
An optimized version of the .head
method without any arguments (as an endpoint), has been implemented.
An optimized version of the .join
method has been implemented.
The .max
method, when called without any named arguments, provides the maximum value without any specific hypering.
The .min
method, when called without any named arguments, provides the minimum value without any specific hypering.
An optimized version of the .minmax
method has been implemented.
An optimized version of the .reduce
method has been implemented.
Caveat: will only produce the correct results if the order of the values is not important to determine the final result.
An optimized version of the .sum
method has been implemented.
An optimized version of the .tail
method without any arguments (as an endpoint), has been implemented.
These methods produce another ParaSeq
object that will keep the parallelization information so that the possibility for parallelization is used whenever possible.
In alphabetical order:
An optimized version of the .antipairs
method has been implemented.
An optimized version of the .batch
method has been implemented.
The nature of the .collate
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .combinations
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
An optimized version of the .deepmap
method has been implemented.
Caveat: to catch any execution of last
, the ParaSeq
module exports its own last
subroutine. This means that any mapper code will either have to live within the scope in which the use ParaSeq
command is executed, to ensure proper last
handling in a hypered .deepmap
.
An optimized version of the .duckmap
method has been implemented.
Caveat: to catch any execution of last
, the ParaSeq
module exports its own last
subroutine. This means that any mapper code will either have to live within the scope in which the use ParaSeq
command is executed, to ensure proper last
handling in a hypered .duckmap
.
Caveat: due to the way duckmap
was implemented before the 2024.06
release of Rakudo, the LEAVE
phaser would get called for every value attempted. This has since been fixed, so that the ENTER
and LEAVE
phasers are only called if the Callable
is actually invoked.
The .flat
method is already very simple, and adding hypering overhead would do just that: add overhead. Therefore, no specific hypering logic has been added for this method.
No specific support for the .flatmap
method has been added, as its core implementation will do the right thing on ParaSeq
objects as well.
An optimized version of the .grep
method has been implemented.
Caveat: if a last
statement is executed, it will ensure that no further values will be delivered. However, this may not stop other threads immediately. So any other phasers, such as ENTER
, and LEAVE
may still get executed, even though the values that were produced, will not be delivered.
Caveat: to catch any execution of last
, the ParaSeq
module exports its own last
subroutine. This means that any matcher code will either have to live within the scope in which the use ParaSeq
command is executed, to ensure proper last
handling in a hypered .grep
.
The .head
method is already very simple, and adding hypering overhead would do just that: add overhead. Therefore, no specific hypering logic has been added for this method.
The nature of the .invert
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The .keys
method is already very simple, and adding hypering overhead would do just that: add overhead. Therefore, no specific hypering logic has been added for this method.
An optimized version of the .kv
method has been implemented.
An optimized version of the .map
method has been implemented.
Caveat: due to a bug in Rakudo until 2024.06 release, any FIRST
phaser will be run at the start of each batch. This can be worked around by using the $, the nameless state variable:
@am.&hyperize.map({ FIRST say "first" unless $++; 42 });
Caveat: if a last
statement is executed, it will ensure that no further values will be delivered. However, this may not stop other threads immediately. So any other phasers, such as ENTER
, LEAVE
and NEXT
may still get executed, even though the values that were produced, will not be delivered.
Caveat: to catch any execution of last
, the ParaSeq
module exports its own last
subroutine. This means that any mapper code will either have to live within the scope in which the use ParaSeq
command is executed, to ensure proper last
handling in a hypered .map
.
An optimized version of the .max
method with named arguments, has been implemented.
An optimized version of the .maxpairs
method has been implemented.
An optimized version of the .min
method with named arguments, has been implemented.
An optimized version of the .minpairs
method has been implemented.
An optimized version of the .nodemap
method has been implemented.
Caveat: to catch any execution of last
, the ParaSeq
module exports its own last
subroutine. This means that any mapper code will either have to live within the scope in which the use ParaSeq
command is executed, to ensure proper last
handling in a hypered .nodemap
.
An optimized version of the .pairs
method has been implemented.
The nature of the pairup
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .permutations
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .pick
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .produce
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .repeated
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .reverse
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .roll
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .rotate
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
An optimized version of the .rotor
method has been implemented for the single argument non-Pair
case. All other cases are basically too complicated to hyper, and therefore have no specific hypering logic.
The simple cases of .skip()
and .skip(N)
are handled by skipping that amount on the result iterator and returning the invocant.
The nature of the other types of arguments on the .skip
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for these cases.
The nature of the .slice
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .snip
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
Since it doesn't make much sense for the .snitch
method to be called on the ParaSeq
object as a whole, calling the .snitch
method instead activates snitcher logic on the ParaSeq
object itself. If set, the snitcher code will be called for each input buffer before being scheduled (in a threadsafe manner). The snitcher code will receive this buffer as a List
.
The nature of the .sort
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
An optimized version of the .squish
method has been implemented if either a :as
named argument has been specified, or a :with
named argument with something other than the ===
operator has been specified. The default case of the .squish
method is already highly optimized so it doesn't make any sense to add any hypering logic for that case.
The nature of the .tail
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
The nature of the .toggle
method basically makes it impossible to hyper. Therefore, no specific hypering logic has been added for this method.
An optimized version of the .unique
method has been implemented.
Caveat: using the :with
option with an operator that checks for inequality, may produce erroneous results.
The .values
is basically a no-op, so it returns the invocant for simplicity.
The following coercers have been optimized as much as possible with the use of &hyperize
. Note that these methods are optimized for speed, rather than memory usage. Should you want to optimize for memory usage, then put in a .Seq
coercer first. Or remove .&hyperize
completely. So:
my $speed = (^1_000_000).&hyperize.Set; # optimized for speed
my $memory = (^1_000_000).&hyperize.Seq.Set; # optimized for memory
my $none = (^1_000_000).Seq.Set; # no parallelization
In alphabetical order:
The Array
coercer has been implemented: it removes all hypering information from its invocant.
Putting together the Bag
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
Putting together the BagHash
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
An optimized version of the Bool
coercer has been implemented.
The .eager
method is functionally the same as the .List
, and has been implemented as such.
Putting together the Hash
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
An optimized version of the Int
coercer has been implemented.
The IterationBuffer
coercer collects the result of the invocant in an IterationBuffer
and returns that.
The list
coercer is implemented: it creates a List
object with the object's iterator as its "todo" (so no elements are reified until they are actually demanded). And thus removes all hypering information.
The List
coercer has been implemented: it creates a fully reified List
and thus removes all hypering information.
Putting together the Map
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
Putting together the Mix
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
Putting together the MixHash
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
An optimized version of the Int
coercer has been implemented.
The Seq
coercer has been implemented: it removes all hypering information from its invocant.
The .serial
method is functionally the same as the .Seq
, and has been implemented as such.
Putting together the Set
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
Putting together the SetHash
of each batch in a correct manner, would be complicated and potentially CPU intensive. Therefore, no specific hypering logic has been added for this coercer at this point.
The Slip
coercer has been implemented: it removes all hypering information from its invocant.
An optimized version of the Str
coercer has been implemented.
The class of which an instance is returned by hyperize
and racify
if the requirements for parallelizing have been met.
Can be used in any situation where a Seq
could also be used. The following additional methods can also be called for introspection and debugging. In alphabetical order:
Bool. Returns whether batch sizes will be automatically optimized to provide the best throughput.
Range. The smallest and largest batch size as a Range
.
Bool. Returns whether exceptions in parallel worker threads will be caught.
say ParaSeq.default-batch; # 16
ParaSeq.default-batch = 1024;
say ParaSeq.default-batch; # 1024
Int. The default initial batch size: currently 16
. Can also be used to change the default batch size by assigning to it.
say ParaSeq.default-degree; # 7
ParaSeq.default-degree = 3;
say ParaSeq.defaultdegreebatch; # 3
Int. The default maximum number of worker threads to be used. Currently set to the number of available CPUs minus one. Can also be used to change the default maximum number of worker threads by assigning to it.
Bag. Returns a Bag
of exceptions that were caught in parallel processing so far, keyed to the .gist
of the execution error (and the value being the number of times they occurred).
By default, any caught exceptions will be printed on STDERR at the end of the process. Calling this method effectively clears the caught exceptions, so that it can also be used to inhibit showing any execution errors.
Change hypering settings on invocant and returns invocant. Takes the same arguments as &hyperize
. Additionally takes a :racing
named argument, to indicate racing
or hypering
semantics.
Bool. Returns whether the ParaSeq
iterator should be considered lazy or not. It will be considered lazy if the source iterator is lazy.
Int. The number of items processed, as obtained from the stats
.
Int. The number of items produced, as obtained from the stats
.
Bool. Returns True
if racing (aka, parallelizing without taking care of the order in which values will be delivered). Else False
.
A List
of ParaStats
objects that were produced, in the order that they were produced. Note that due to the asynchronous nature of stats production and processing, this may be incomplete at any given time.
Returns a List
of the IDs of the CPU threads that have been used to produce the result.
The ParaStats
class is created for each batch that has been processed. It provides the following methods (in alphabetical order):
Int. The number of nano-seconds that it took to process all values in the associated batch.
Int. The ordinal number of the associated batch (starting at 0).
Int. The number of values processed in the associated batch (aka the batch size of that batch).
Int. The number of values that were produced in the associated batch. This can be 0
, or a higher value than processed
when .map
was called with a Callable
that produced Slip
values.
Int. The number of nano-seconds that it took to process all values in the associated batch, corrected for any Garbage Collection cycles.
Int. The numerical ID of low-level thread that processed the associated batch (aka its $*THREAD.id
).
One of the main issues with the current implemementation of .hyper
and .race
in the Rakudo core is that the batch size is fixed. Worse, there is no way to dynamically adapt the batch size depending on the load.
Batch sizes that are too big, have a tendency to not use all of the CPUs (because they have a tendency to eat all of the source items too soon, thus removing the chance to start up more threads).
Batch sizes that are too small, have a tendency to have their resource usage drowned out by the overhead of batching and dispatching to threads.
This implementation automatically adapts batch sizes from the originally (implicitely) specified batch size for better throughput and resource usage. If this feature is not desired, the :!auto
argument can be used to switch automatic batch size adjustment off.
If the degree
specified is 1, then there is no point in batching or parallelization. In that case, this implementation will take itself completely out of the flow.
Alternately, if the initial batch size is large enough to exhaust the source, it is clearly too large. Which is interpreted as not making any sense at parallelization either. So it won't.
Note that the default initial batch size is 16, rather than 64 in the current implementation of .hyper
and .race
, making the chance smaller that parallelization is abandoned too soon.
The .serial
method or .Seq
coercer can be typically be used to "unhyper" a hypered sequence. However many other interface methods do the same in the current implementation of .hyper
and .race
, thereby giving the impression that the flow is still parallelized. When in fact they aren't anymore.
Also, hyperized sequences in the current implementation are considered to be non-lazy, even if the source is lazy.
This implementation aims to make all interface methods pass on the hypered nature and laziness of the sequence.
Some loop control statements may affect the final result. Specifically the last
statement does. In the current implementation of .hyper
and .race
, this will only affect the batch in which it occurs.
This implementation aims to make last
stop any processing of current and not create anymore batches.
Currently only the .map
and .grep
methods are supported by the current implementation of .hyper
and .race
(and not even completely). Complete support for .map
and .grep
are provided, and other methods (such as .first
) will also be supported.
When an interface method takes a Callable
, then that Callable
can contain phasers that may need to be called (or not called) depending on the situation. The current implementation of .hyper
and .race
do not allow phasers at all.
This implementation aims to support phasers in a sensible manner:
Called before each iteration.
Called on the first iteration in the first batch.
Called at the end of each iteration.
Called on the last iteration in the last batch. Note that this can be short-circuited with a last
control statement.
Called after each iteration.
One of the great features of the Raku Programming Language, is the handling of regexes (aka regular expressions). However, in a multi-threaded environment, one of its features: the $/
is a potential source of problems. That is because if a block of code doesn't define its own $/
variable, the nearest lexically scoped $/ will be used.
For example:
say @words.&hyperize.grep({ / foo / }).elems;
has a great chance of not producing the correct results, especially for larger number of elements in @words
. That's because all of the threads running the code inside the .grep
will share the nearest lexically visible $/
. In Rakudo releases before 2024.08, it could even cause crashes. Since then, it will only produce potentially incorrect results.
The fix is pretty easy: make sure a my $/;
is defined inside the scope of the block. Taking the above example:
say @words.&hyperize.grep({ my $/; / foo / }).elems;
A description of the program logic when you call .&hyperize
on an object.
If the degree is 1, then there's nothing to parallelize. The &hyperize
sub will return its first argument. Since .&hyperize
will most likely mostly be called as a method, the first argument will be described as the "invocant".
Obtain an iterator from the invocant and read as many items into an IterationBuffer
as the initial batch size indicates. If this exhausted the source iterator, then return a Seq
on that IterationBuffer
. This also effectively inhibits any parallelization.
If step 1 and 2 didn't result in a premature return, then a ParaSeq
object is created and returned. Apart from the initial IterationBuffer
and source iterator, it also contains:
-
the initial batch size
-
the degree (max number of worker threads)
-
the "auto" flag
-
the $*SCHEDULER value
Each interface method decides whether they can support an optimized hypered operation. If does not support hypered operation, a special BufferIterator
is created. This iterator takes the buffer and the source iterator from step 2, that will first deliver the values from the buffer, and then the values from the source iterator (until that is exhausted).
The following step then depends on whether the method is an endpoint (such as max
) or a coercer (such as Str
). If it is, then the BufferIterator
will be wrapped in a standard Seq
object and the core equivalent of that method will be called on it and its result returned.
If it is an other interface method, then a clone of the original invocant is created with the BufferIterator
object as its source iterator. And that is then returned.
If the method can be parallized, then we continue.
Some methods (such as map
and grep
) may require more than 1 value per iteration (the granularity). Batches need to have a size with the correct granularity, otherwise processing a batch may fail at the last iteration for lack of positional arguments.
So read additional values from the source iterator until the granularity is correct.
One of the issues with hypering, is that multiple threads can easily produce so many values that the deliverer can be overwhelmed in the worst case. Or at least, it could be wasteful in resources because it would produce way more values than would actually need to be delivered. For example:
say (^Inf).&hyperize.map(*.is-prime)[999999]; # the millionth prime number
One would not want the parallelization to calculate 1.2 million prime numbers while only needing 1 million. Yet, this will happen quite easily if there is not enough "back pressure".
So the thread delivering values, needs to be able to tell the batcher when it is ok to schedule another batch for processing. This is done through a ParaQueue
object (a low-level blocking concurrent queue to which the delivering thread will push whenever it has delivered the results of a result batch. Since "back pressure queue" is a bit of a mouthful, this is being referred to as the "pressure queue".
The deliverer is also responsible for telling the batcher what it thinks the best size for the next batch is. See Step 6.
To get things started, a pressure queue is set up with various batch sizes (one for each degree) so that maximally that many threads can be started. By using variable sizes, the delivery iterator can more quickly decide what the optimum size is for the given workload because the different batch sizes produce different average time needed per processed value. Which is used to calculate the optimum batch size.
The result queue is where all of the worker threads post the result of processing their batch. The result queue is a ParaQueue
of ParaQueue
s. Whenever the batcher is told that it may queue up the next batch, the batcher also pushes a new ParaQueue
object to the result queue and passes that ParaQueue
object to the worker thread. So that when the worker thread is done processing, it can push its result to that ParaQueue
.
The ParaQueue
that is pushed to the result queue by the batcher ensures that the order in which batches of work are handed out, will also be the order in which the results will be delivered. The fact that the result of processing is pushed to this ParaQueue
itself, makes the deliverer block until the worker thread has pushed its result (even if other threads may have been ready earlier).
To communicate back performance statistics from a worker thread finishing their work, another ParaQueue
is created. Its sole purpose is to collect all ParaStats
objects that each worker thread has completed as soon as they are done. It basically also represents the number of result batches that are waiting to be delivered: hence the name "waiting queue".
This queue is inspected by the delivery thread whenever it has completed delivery of a batch. It's possible that this queue is empty. It can also contain the statistics of more than one worker thread.
The batcher thread keeps running until the source iterator is exhausted. When it is exhausted, it pushes an IterationEnd
to the result queue, indicating that no more batches will be coming. It's the only time a non-ParaQueue
object is pushed to the result queue.
The batcher is basically a loop that blocks on the messages from the deliverer. Whenever a message is received (with the proposed batch size, calculated from any available performance statistics), it adjusts for correct granularity (as the deliverer is unaware of any granularity restrictions). The batcher then fetches that many values from the source iterator and schedules another job for a worker thread with the values obtained (if any).
At this point a ParaIterator
object is created and installed as the (result) .iterator
of the invocant. The paralellizing logic in the hypered variant of the method is started, which then in turn starts feeding the ParaIterator
.
It is then up to the code that iterators over the (result) iterator to consume the values delivered. If no values are being delivered, production of values will halt after all inital "degree" batches have been produced.
When the deliverer is done with a result batch, it removes all available statistics information and adds these to the stats
of the ParaSeq
object. It also uses this information to calculate an optimum batch size for any other batches to be processed (unless this is inhibited by the :!auto
named argument to .&hyperize
).
The algorithm for batch size calculation is pretty simple at the moment. To allow for optimum responsiveness of the program, as well as any other processes running on the computer, a time slice of 1_000_000 nanoseconds (aka 1 millisecond) has been chosen.
If a batch took almost exactly that amount of time to produce values, then the size of that batch is assumed to be the optimum size. If the batch took longer, then the batch size should be reduced. If the batch took less than that to complete, then the batch size should be increased.
source
\- hyperize
|- invocant (degree == 1)
|- Seq (source exhausted with initial batch)
|- ParaSeq
\- method call
|- endpoint
|- coercer
|- hypering
/---------------------------------------------------------\
| | |
| ⬇ |
| -------------- ------------ |
| | statistics | /-----⮕ | producer | ⮕ -|
| -------------- / ------------ |
⬇ ⬇ / |
----------- ------------ ----------- ------------ |
| results | ⮕ | deliverer | ⮕ | batcher | ⮕ | producer | ⮕ -|
----------- ------------ ----------- ------------ |
| \ |
| \ ------------ |
| \-----⮕ | producer | ⮕ -/
⬇ ------------
values
consumed by program
Elizabeth Mattijsen [email protected]
Source can be located at: https://github.com/lizmat/ParaSeq . Comments and Pull Requests are welcome.
If you like this module, or what I’m doing more generally, committing to a small sponsorship would mean a great deal to me!
Copyright 2024 Elizabeth Mattijsen
This library is free software; you can redistribute it and/or modify it under the Artistic License 2.0.