AtomicObjects

Usage

use AtomicObjects;

or

import AtomicObjects;

Support for atomic operations on pointers to ‘unmanaged’ classes.

Warning

This module has several platform restrictions in its current state:

  • It relies on Chapel extern code blocks and so requires that the Chapel compiler is built with LLVM enabled.

  • The implementation relies on using either GCC style inline assembly (for x86-64) or a GCC/clang builtin, and so is restricted to a CHPL_TARGET_COMPILER value of gnu, clang, or llvm.

  • The implementation does not work with CHPL_ATOMICS=locks.

This module provides support for performing atomic operations on pointers to unmanaged classes, which can be thought of as building blocks for creating non-blocking algorithms and data structures.

Note

Only unmanaged classes are supported as they are represented, internally, as raw 64-bit or 128-bit pointers. We do not support shared or owned objects as they are essentially wrappers around an unmanaged object, and borrowed objects require a static lifetime.

Global Atomics

By default, the AtomicObject can support atomic operations on potentially remote objects. This does add some additional overhead and can be turned off during initialization.

var atomicVar : AtomicObject(unmanaged Obj, hasGlobalSupport=false);

Warning

Currently, AtomicObject only supports up to 65535 locales, and also works on the assumption that only the lowest 48 bits of the virtual address space will ever be used. An exascale solution that allows for an arbitrary number of compute nodes and for the entire 64-bit address space to be utilized is future work that is in-progress.

Note

When hasGlobalSupport=true and hasABASupport=false, it will enable RDMA atomics, I.E when CHPL_NETWORK_ATOMICS!="none", which is provides a significant improvement in performance on systems where they are support, notable on a Cray-XC.

Warning

Currently, hasGlobalSupport=true is necessary when using it from multiple locales, even if it is intended to be used locally. This is due to there being no compiler primitive to create a ‘wide’ class, nor a way to cast a wide-pointer to create a wide class.

ABA Wrapper

The ‘ABA’ problem occurs when a task T1 reads the value A from location L, another task T2 writes B to L, and another task T3 writes the value A to L; once T1 checks to see if L has changed, it will incorrectly assume that it has not. To make this more concrete, think of A and B both as a node in a linked list; T1 reads A, T2 allocates a new node B and writes it to L and deletes A, and T3 allocates a new node which just so happens to be the same piece of memory that A had before and writes it to L. Atomic operations such the compareAndSwap will succeed despite the fact that the nodes are not the same as it will perform the operation based on the virtual address.

The ABA wrapper is one solution to this problem by coupling a 64-bit count alongside the normal 64-bit virtual address or the 48-bits of virtual address and 16-bit locality information. AtomicObject has its own ABA variants of its API, which can both take and return ABA wrappers. Examples of how they can be used can be observed below. It is safe to mix-and-match both ABA and non-ABA variants of the API, but only the ABA variants will advance the ABA counter.

var atomicVar : AtomicObject(unmanaged Obj, hasABASupport=true);
var obj1 = new unmanaged Obj();
var obj2 = new unmanaged Obj();
atomicVar.write(obj1);
var a = atomicVar.readABA();
var b = atomicVar.writeABA(obj2);
atomicVar.writeABA(obj1);
assert(atomicVar.compareAndSwap(obj1, obj2) == false, "This should always fail!");

Note

We forward all accesses to the ABA wrapper to the object it is wrapping so that whether or not the ABA versions of the AtomicObject API is used, it becomes as transparent as possible. This applies to all method and field accesses.

record ABA : serializable

Wrapper for an object protected by an ABA counter. This type forwards to the object represented by its underlying pointer and hence can be used as if it were the object itself, via ‘forwarding’. This type should not be created by the user, and instead should be created by LocalAtomicObject. The object protected by this ABA wrapper can be extracted via ‘getObject’.

type __ABA_objType
proc init(type __ABA_objType)
proc init=(other: ABA(?objType))
proc getObject() : __ABA_objType?
proc getABACount()
proc serialize(writer, ref serializer) throws

Writes an ABA

operator ABA.=(ref lhs: ABA, const ref rhs: lhs.type)
operator ABA.==(const ref aba1: ABA, const ref aba2: ABA)

Special case operator that compares two ABA wrappers.

operator ABA.!=(const ref aba1: ABA, const ref aba2: ABA)
record AtomicObject : serializable
type objType
param hasABASupport : bool
param hasGlobalSupport : bool
var atomicVar : if hasABASupport then _ddata(_ABAInternal(objType?)) else atomic uint(64)
proc init(type objType, param hasABASupport = false, param hasGlobalSupport = !compiledForSingleLocale())
proc init(type objType, defaultValue: objType, param hasABASupport = false, param hasGlobalSupport = !compiledForSingleLocale())
proc init=(other: AtomicObject)
proc readABA() : ABA(objType?)
proc ref read() : objType?
proc ref compareAndSwap(expectedObj: objType?, newObj: objType?) : bool
proc compareAndSwapABA(expectedObj: ABA(objType?), newObj: objType?) : bool
proc compareAndSwapABA(expectedObj: ABA(objType?), newObj: ABA(objType?)) : bool
proc ref write(newObj: objType?)
proc write(newObj: ABA(objType?))
proc writeABA(newObj: ABA(objType?))
proc writeABA(newObj: objType?)
proc ref exchange(newObj: objType?) : objType?
proc exchangeABA(newObj: objType?) : ABA(objType?)
proc exchangeABA(newObj: ABA(objType?)) : ABA(objType?)