nim-works / loony

A high throughput MPMC lock-free queue based on a paper by Giersch, Nolte et al implemented in pure Nim.

nim-works

loony's Issues

problem with strings, something I should know about?

Hi, I'm quite new to Nim and I work on it occasionally during weekends.

I'm testing loony for a simple double thread producer/consumer paradigm. Here's the code:

import std/[locks, os]
import loony

type
  Message = object
    value: string

let fifo:LoonyQueue[Message] = newLoonyQueue[Message]()
var L: Lock
var producing = true

proc producer() {.thread.} =
  for i in 1..10:
    let Message = Message(value: "Message " & $i)
    acquire(L)
    fifo.push Message
    release(L)
    echo "Produced: ", Message
    sleep(100)

proc consumer() {.thread.} =
  while producing:
    acquire(L)
    let item = fifo.pop
    if item.value != "":
      echo "Consumed: ", item
    release(L)
    sleep(10)

# Create worker threads
var producerThread, consumerThread: Thread[void]

initLock(L)

# Start worker threads
createThread(producerThread, producer)
createThread(consumerThread, consumer)

joinThread(producerThread)
producing = false
joinThread(consumerThread)

deinitLock(L)

It's a simple, documentation-alike, piece of software, but when I run it I get this output:

Produced: (value: "Message 1")
Produced: (value: "Message 2")
Produced: (value: "Message 3")
Consumed: (value: "��\x0Ew\x14�\x03@")
Produced: (value: "Message 4")
Produced: (value: "Message 5")
Consumed: (value: "��\x0Ew\x14�\x03@")
Produced: (value: "Message 6")
Produced: (value: "Message 7")
Consumed: (value: "��\x0Ew\x14�\x03@")
Produced: (value: "Message 8")
Produced: (value: "Message 9")
Consumed: (value: "��\x0Ew\x14�\x03@")
Produced: (value: "Message 10")
Consumed: (value: "��\x0Ew\x14�\x03@")

I would like to talk about the weird codes I get from the popped Messages' values and asking if there's something I don't know about strings that explain this.
I also tested the same code using a LoonyQueue[string] directly, and in this case the code didn't compile at all

Using MacOS Catalina 10.15.7, nim 1.6.12, latest loony

Example code errors

Hi, I tried to use the example code from the readme (https://github.com/nim-works/loony#how-to-use) but I got some errors

Hint: used config file '/home/---/.choosenim/toolchains/nim-#devel/config/nim.cfg' [Conf] Hint: used config file '/home/---/.choosenim/toolchains/nim-#devel/config/config.nims' [Conf] ................................................................ /home/---/Desktop/Nim/test.nim(10, 11) template/generic instantiation ofpushfrom here /home/---/.nimble/pkgs/loony-0.1.1/loony.nim(177, 27) template/generic instantiation offetchIncTailfrom here /home/---/.nimble/pkgs/loony-0.1.1/loony.nim(71, 26) Error: attempting to call undeclared routine: 'fetchAdd'

This happens both on the 0.1.0 and 0.1.1

Introduce balls

Make tests with balls.

• Test single thread low/high volume push and pop ensuring exact retrieval and reallocation of nodes
• Test threaded (2) independent high volume producer and consumer
• Test threaded (2) mixed producer and consumer
• Create suite testing incremental numbers of threads with mixed producer and consumer qualities

thread sanitizer does not agree with loony

ico@workdoos:/tmp/loony(main)$ nim r  --passC:-fsanitize=thread --passL:-fsanitize=thread --threads:on tests/test.nim 
[...]
🟢 creation and initialization of the queue
FATAL: ThreadSanitizer CHECK failed: ../../../../src/libsanitizer/tsan/tsan_interface_atomic.cpp:223 "((IsLoadOrder(mo))) != (0)" (0x0, 0x0)
    #0 __tsan::TsanCheckFailed(char const*, int, char const*, unsigned long long, unsigned long long) ../../../../src/libsanitizer/tsan/tsan_rtl_report.cpp:47 (libtsan.so.0+0x972d2)
    #1 __sanitizer::CheckFailed(char const*, int, char const*, unsigned long long, unsigned long long) ../../../../src/libsanitizer/sanitizer_common/sanitizer_termination.cpp:78 (libtsan.so.0+0xb513a)
    #2 AtomicLoad<long long unsigned int> ../../../../src/libsanitizer/tsan/tsan_interface_atomic.cpp:223 (libtsan.so.0+0x762e4)
    #3 __tsan_atomic64_load ../../../../src/libsanitizer/tsan/tsan_interface_atomic.cpp:539 (libtsan.so.0+0x762e4)
    #4 load_OOZloonyZnode_116 <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x6baa1)
    #5 advTail_test_334 <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x72289)
    #6 push_test_306 <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x731f4)
    #7 enqueue_test_303 <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x73378)
    #8 NimMainModule <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x819d5)
    #9 NimMainInner <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x797f3)
    #10 NimMain <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x79849)
    #11 main <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x798b9)
    #12 __libc_start_main ../csu/libc-start.c:308 (libc.so.6+0x26d09)
    #13 _start <null> (test_EF21B045952947A3759AD00DE01B49BBBC3590C4+0x94d9)

Error: execution of an external program failed: '/home/ico/.cache/nim/test_d/test_EF21B045952947A3759AD00DE01B49BBBC3590C4 '

introduce queue state?

I'd like to have some controls on the queue:

• I want to be able to pause the queue for pop clients (so they receive nil).
• I want to have a way to clear the entire queue, discarding the contents.
• I want a backchannel for pop that works while the queue is paused so that I can, for example, tear down items in the queue with extra logic.

Now, maybe the right way to implement this is with a queue manager that sits outside the queue itself, but in any event, these are pretty minor features that we should be able to support for programmer comfort.

[MacOS/Darwin] Emulating Futex behaviour

Issue

Currently I use linux futex and windows waitonaddress to manage backpressure when using Wards and Loony as a threadpool queue.

Mac/Darwin do not have primitives or APIs that are similar.

Solutions

At the moment, I can only think of using conditions and locks.

[NIM] Thread completion triggers thread sanitizer warnings

Thread analyzer determines the deallocshared proc called by a thread on completion is a data-race.

Atm only solution I can come up with involves changing or providing a modified std lib thread library that corrects this behaviour. The safest methodology is having the pointer deallocshared acts on being atomically stored and loaded for destruction? Haven't look too deep into it yet.

Related to #9 and #13

Fail to deallocate consumed nodes

Nodes aren't hitting their memory reclamation operations

Related stuff worth looking at before implementation :)

a) Nim's Weave (which offers different queues, work-stealing scheduler with some twists, but most importantly excellent performance - I don't know of any other as generic solution which scales nearly linearly and uses so high percentage of computing power per CPU)

b) https://github.com/ConorWilliams/Forkpool

[CHORE] Refactor alias types

Refactor all these alias types like TagPtr to use a distinct type with converters to avoid boilerplate spaghetti as per disruptek

SIGSEGV: Illegal storage access. (Attempt to read from nil?)

Using a slightly modified example code for more consumer threads my system throws SIGSEGV error.

import std/[locks, os, atomics]
import loony

type
  Message = ref object
    value: string

let fifo = newLoonyQueue[Message]()
var terminate: Atomic[bool]


proc producer() {.thread.} =
  var n: int = 0
  for i in 1..1000000:
    n.inc()
    let msg = Message(value: "Message " & $i)
    if n > 10000:
      echo "Producing ", repr(msg)
      n = 0
    fifo.push msg
    #sleep(0)
  terminate.store(true)

proc consumer() {.thread.} =
  var n: int = 0
  while true:
    let item = fifo.pop
    if not item.isNil:
        n.inc()
        if n > 10000:
            echo "Consumed: ", repr(item)
            n = 0
    else:
        if terminate.load:
            break
    #sleep(0)

# Create worker threads
var producerThread: Thread[void]
var consumerthreads: array[512, Thread[void]]

const THREADS: int = 14

# Start worker threads
createThread(producerThread, producer)
for t in 0..<THREADS:
    echo "create thread", t
    consumerthreads[t].createThread(consumer)

joinThread(producerThread)

for t in 0..<THREADS:    
    joinThread(consumerthreads[t])

SIGSEGV on q.isEmpty

Hi,

i see SIGSEGV on calls to queue.isEmpty. The queue is initialized using newLoonyQueue. It has not received any items yet. I tried if- and while - makes no difference. I tried std/atomics and threading/atomics. It feels a bit like a timing/init-problem. Though i never managed to make it work reliably. Goes like this - the whole app runs thru and finishes. Then on the next run it breaks early. On the following start i might run and finish or break in-between.
I use the main-thread to fill the queue and one separate thread to read and process the items from the queue ? A queue-items contains a seq[int], int, string, so nothing special here.

Im on Nim 1.6.6 / macOS

this is my thread-proc ::

proc warden[K, V]( map: Map[K, V] ) {.thread, nimcall.} =

  echo "maintanance:: started"

  while true:

    while map.workQ.isEmpty:
      sleep 500
      continue

    var elem = map.workQ.pop
    #if elem.isNil: break
    echo &"ward :: ", elem.task, " :: ", elem.parent, " - ",elem.slot

    if elem.slot == -1:
      echo "\nmaintanance:: shuts down..\n"
      return

Traceback for :

nim r -d:b32 --gc:arc --threads:on -d:useMalloc -d:debug -d:nimOldCaseObjects Mapi.nim

/Users/asc/sw/nim/ctrie/champ/Mapi.nim(30) warden
/Users/asc/.nimble/pkgs/loony-0.1.12/loony.nim(68) isEmpty
/Users/asc/.choosenim/toolchains/nim-1.6.6/lib/pure/concurrency/atomics.nim(336) load
SIGSEGV: Illegal storage access. (Attempt to read from nil?)
Error: execution of an external program failed: '/Users/asc/.cache/nim/Mapi_d/Mapi_558E24AF8FBA3F576C55E835091AA4F6A61AFFB3 '

Debugging helpers

Using some kind of flag like -d:loonyDebug want to be able to count and relay the number of enqueues and dequeues performed with the use of a queue; allocations and deallocations of nodes; index tracking/alert when reaching overflow etc. I don't mind if it affects performance.

[LOONY] Memory Leak 0.1.3

As with all great things, somewhere along the way we screwed up. The nodes memory reclamation algorithm is not correct for some changes that were made while refactoring.

Currently the leak has been corrected up to a degree of ~+90% (with previous test node leaks of about 29 now down to 2+/-2). I’m on the last stretch of correcting the remainder.

you can see #15 for progress

[ORC] Crash when GC attempts to collect element

Orc crashes when a Continuation (or any ref object) passed through the queue reaches the end of its life span and the final =destroy is run.

This is evident by incrementing the ref count of an object by 1 before the end of its lifespan and comparing the behaviour before and after.

ARC does not share this issue; the final =destroy can be called without issue and the memory is collected appropriately.