[LBNL]

Berkeley UPC - Unified Parallel C

(A joint project of LBNL and UC Berkeley)
[UCB]

Home
Downloads
Documentation
Bugs
Publications
Demos
Contact
Internal

Memory Management in the UPC Runtime

Version 1.2, March 31, 2013

Paul H. Hargrove < >

This is a design document for the implementation of memory management in the runtime layer of the Berkeley UPC compiler. The external API for the runtime layer (as seen by compiler-generated code) is documented in the UPC Runtime Specification.


Notational conventions

Prefix Meaning
upc_ indicates a UPC language-level symbol
upcr_ indicates a symbol in the runtime system not visible to the UPC programmer (called by compiler-generated code)
upcri_ indicates a symbol in the runtime system not visible outside the runtime system (called only within the runtime)
gasnet_ indicates a routine in the GASNet communication system (not visible to users or compiler: called by runtime functions)

"Threads" vs. Pthreads/Processes

"Thread" will be used throughout this document to indicate a parallel thread of execution in the UPC job. Where processes vs. pthreads need to be treated differently for implementation reasons, context will make it clear which we are talking about.

"Static" User data

In this document, 'static' user data means "not dynamically allocated" (as opposed to static vs. global variable declaration). Both global variables and those declared with the 'static' keyword (at either file or function scope) are 'static' in this sense.


Supported Platform Environments, by Type of Memory

The UPC runtime will support 3 different runtime environments, which differ in the way that accesses to (logically) shared memory between UPC threads are done. The type of runtime will be determined at compile time, and will cause the UPCR_PLATFORM_ENVIRONMENT preprocessor macro to be set to one of the following values:
  1. UPCR_PURE_SHARED

    On pure shared memory machines, shared memory (and interprocess mechanisms such as pthreads/SysV IPC synchronization) is used to perform all interthread communication.

  2. UPCR_PURE_DISTRIBUTED

    In this model, the UPC runtime will reply upon the network (i.e. the GASNet layer) for all inter-thread communications.

  3. UPCR_SHARED_DISTRIBUTED

    On CLUMPs ("clusters of SMPs"), multiple threads/processes will exist on each of multiple nodes. In this environment, shared memory will be used for communication between pthreads/processes on the same node, and GASNet will be used for communication between nodes.


Memory Layout of a UPC Process

UPC Process Memory Map The diagram on the right shows a possible layout for the memory regions of a single process that is implementing 2 UPC threads (via 2 pthreads): the sections need not necessarily be in the order specified. A UPC that only implements a single UPC thread would not use pthreads, and would thus only have 1 stack, 1 global and local shared heap, and no thread-local data sections.

Stack, program text, and program data

The program text, data, and stack of a UPC executable will be laid out and managed in the traditional fashion by the linker and loader, with the text and data toward the start of the address range, and the stack at the top. If pthreads are used, the pthreads library will take care of splitting the user stack into separate, per-thread stacks (we may wish to allow users to specify the size of the thread-specific stacks, in case they need to use larger stacks than the pthreads' library's default size).

Thread-local data

Any unshared global or static data declared in a UPC program is "thread-local", i.e. if pthreads are used, a separate copy for each pthread must be allocated, and uses of the data must be handled specially so that the correct copy of the data is used (see the implementation note on handling static user data for more details). The diagram here assumes we are using the 'tld section' approach to thread-local data, in which we place all such data in a special section in the executable, then make copies of it for new threads (the original program thread--thread 0--can use the original section). In the alternative 'global struct' approach, each thread would store its data in a large C struct allocated out of the regular C heap.

When pthreads are not used, the user's static/global unshared variables simply exist as regular C static/global variables in the program data section.

Shared address space

As can be seen from the memory layout diagram, all shared memory will be allocated within a single, large contiguous address space. The size and limits of this region will be fixed at startup time (as with the local heap, users will be able to specify the size of the region). If there are multiple threads, the shared memory area will be split between them, with each thread getting a equal, page aligned slice of the shared address space. Making the address spaces page-aligned and contiguous with one another will generally make shared pointer calculations easier and faster (note: since shared pointer implementations are interchangeable at library compilation time, it will be possible for a slower implementation to be written to handle non-contiguous regions, without affecting performance when a contiguous region is used).

The shared memory region needs to be allocated specially for many network interfaces (the pages may need to be pinned in memory, and/or DMA registration may need to occur, etc.), and so the entire region is allocated for the runtime by GASNet as part of the gasnet_attach() call. GASNet may or may not allocate physical memory for the shared region (it may just mmap some virtual memory to be allocated on first write). It must be guarantee, however, that the memory area returned is legal to reference at any time after gasnet_attach without causing a segmentation fault (this constraint considerably reduces the communication requirements needed for the UPC global shared memory allocation functions--see below).

Shared libraries

UPC programs will generally be compiled statically: however, there may be cases in which dynamic shared libraries will be loaded, and the location where they are put may not be under the control of the UPC runtime or GASNet. Most operating systems have a standard address at which they start performing the mmaps of shared library contents (the x86 Linux loader, for instance, starts loading shared objects starting at address 0x40000000, leaving wide gaps between the loaded libraries' address spaces and those of the stack and default heap). Depending on the amount of shared memory that is needed in a program, this default location may make it difficult to find a large enough contiguous region in which to place the shared memory region (particularly on 32 bit machines).

Dealing with this issue is left to GASNet: the runtime merely requests a shared region of a specific size (with a guaranteed heap offset, as discussed next), and GASNet either provides it or returns an error.

Regular heap

The regular C heap cannot be allowed to overlap into the shared memory region. In the worst case scenario, on some platforms malloc(), etc., may need to be intercepted, and a custom allocator used to guarantee this. However, on most platforms the regular heap will not encroach on any region that has already been mmapped, so the shared memory region will be safe from interference. When the standard C library's heap hits an mmapped area, sbrk() will fail, and the heap will either stop growing (and return NULL when additional memory would be needed to satisfy a malloc()), or grow non-contiguously (i.e. via mmap instead of sbrk).

To support malloc implementations that simply fail once sbrk ceases to provide more memory space, the gasnet_attach function will take a parameter that guarantees a certain spacing between the shared heap area and the beginning of the regular heap. This will guarantee that the shared region (at least) will not limit the size of the regular heap. Users will be able to specify the size of this region via environment variables and/or command line flags.


Categories of Dynamic Memory

As can be seen from the diagram above, each UPC thread will make use of 3 separate memory heaps. If a UPC process uses pthreads, the pthreads will share a single regular heap, but will each have separate shared global and shared local heaps. The following table summarizes which types of allocations reside in which heap:

Heap Types of Allocations Example
Global shared heap UPC shared allocations that result in phased memory blocks that stripe across nodes Any call to upc_global_alloc or upc_all_alloc in which the block size is less than the size of the allocated memory (i.e. any such memory that doesn't exist only on thread 0). Static UPC arrays with the same property.
Local shared heap UPC shared memory that is allocated to exist only on a single node/thread All calls to upc_alloc. Memory allocated by upc_all_alloc and/or upc_global_alloc that exists entirely on node 0 (i.e. has a block size larger than the allocated size). Static shared data with the same property.
Regular heap Unshared memory from standard calls to malloc(), etc. Malloc, calloc, realloc, etc.

Note that since static shared data will be implemented using dynamic allocations at startup--as explained in the notes on static user data--it will exist in the heaps, rather than in a regular data segment.


Use of 3rd Party Memory Allocator(s)

The UPC runtime will provide an interface for memory allocators, so that different implementations may be used when this is desirable (as is may be for debugging, profiling, or performance purposes).

Constraints on Memory Allocators

Several important constraints exist for a memory allocator to be compatible with the UPC runtime:

Interface to Memory Allocators

We will provide separate interfaces to the 3 types of allocators. The preprocessor (or inline functions) can be used to translate these generic functions to the calls provided by a specific allocator (i.e. the addresses of these calls will never be used).

In the functions below <AT> stands for the allocator type, which may be one of: 'sharedglobal', 'sharedlocal', or 'local'. upcra_ stands for UPC Runtime Allocator API, a different name since this is not intended to be called by compiler-generated code, which will use only the global_alloc, etc., style functions.

  upcra_<AT>_heap_t* upcra_<AT>_init(void *heapstart, uintptr_t initbytes);
Called by upcr once during startup (during upcr_init) to initialize each allocator before any other calls to the allocator. heapstart points to the start of the allocator's heap, and initbytes provides an initial size for the heap. The return value is a pointer to a heap structure, which is passed to the other functions to specify which heap is to be used.
  void *upcra_<AT>_alloc(upcra_<AT>_heap_t* heap, size_t numbytes);
Allocates a region of numbytes bytes from the appropriate heap and return a pointer to the allocated memory, which is suitably aligned for any kind of variable. The memory is not cleared.

Returns NULL to indicate that the allocator has run out of free memory and either some memory frees must take place or the allocator needs to be provided with additional pages (using upcra_provide_pages_<AT>()) before a request for numbytes can be satisfied.

  void upcra_<AT>_free(upcra_<AT>_heap_t* heap, void *ptr);
Free the region of memory indicated by ptr, which was previously returned by a call to upcra_alloc_<AT> on this same allocator. Allocators that optionally choose to detect deallocation errors (e.g. duplicate free calls) should print an error message to stderr and call abort() when an error is detected.
  void upcra_<AT>_provide_pages(upcra_<AT>_heap_t* heap, uintptr_t numbytes);
The runtime may call this function at any time to provide additional unused memory pages to the allocator. The allocator will be told it has an additional amount of 'numpages' pages of memory to use (the amount provided will always be a multiple of the system's page size). This new memory will start from the previous end address of the heap (if the global heap grows down, the new memory comes from the bottom end of the heap).

This function will often be called after a failed call to upcra_alloc_<AT> which indicates the allocator requires more pages.

The runtime is responsible for ensuring that the address ranges given to the heaps are safe to access (and, for the shared heaps, have been registered with gasnet such that they can be accessed by remote nodes).

Finally, if the allocator for the global shared heap cannot support growing the address range for the heap downward instead of upward, UPCR_GLOBALHEAP_GROWS_UP must be #defined, as the logic for detecting global/local heap collision is different in that case.


GASNet Critical Sections and Asynchronous Handler Execution

To handle memory allocation requests that arrive from other nodes, the UPC runtime registers 'handler' functions for various message types with the GASNet network layer. When a network message containing a request from another node arrives, it will generally be handled synchronously, as periodic polling is done to see if new messages have arrived. Such polling will happen at most of the UPC Runtime API's entry points--every time a message is to be sent, or a barrier entered, etc.

However, since user code may take an arbitrary amount of time in between calls to the UPC runtime, it is desirable to also have a method of asynchronously forcing incoming requests to be handled, if they have not been handled after a specified amount of time. The GASNet layer may thus cause a signal to be sent (presumably a SIGALRM), in order to force progress on requests. This may result in the handlers being run within a signal handler context.

Traditionally, code running in signal handler context is extremely circumscribed in what it can do: none of the standard pthreads/System V synchronization calls are on the list of signal-safe functions (for such a list see Richard Stevens' *Advanced Programming in the Unix Environment*, p 278). To overcome these limitations, and allow our handlers to be more useful, the normal limitations on signal handlers will be avoided through the use of "Handle-safe locks." All function calls that are not signal-safe and which are used within our handlers MUST be called within a GASNet critical section:

    GASNET_CRITICAL_START
    ...
    GASNET_CRITICAL_END
Critical sections are designed to be held briefly (no blocking, no network calls, etc.). They guarantee that the code within the section will not be interrupted by a signal-spawned GASNet-registered handler (if a signal arrives while the critical section is in effect, the handler will be postponed until the critical section is completed). This system guarantees that 'signal unsafe' (i.e. non-reentrant) functions will not be called reentrantly, and so we may use them even in handler code that may run in signal handler context.

Code that runs inside an asynchronous handler is implicitly within a GASNet critical section, so the START/END macro calls must not be used. Code that runs in regular context (including the user's UPC code) must use these macros whenever a signal-unsafe function is called. This means that the UPC runtime must intercept any such calls (like malloc. etc.), and redirect them to a version that wraps the standard library call with the GASNet critical section macros.

See the GASNet specification for details.

Shared Memory Allocation Algorithms

The pluggable memory allocators used by the UPC Runtime will handle the traditional issues of heap management: allocating and freeing blocks of memory while avoiding fragmentation. The UPC runtime will delimit the memory space controlled by each allocator, handle the internode communication needed for shared memory allocations, and perform most of the synchronization needed around allocations.

The issues involved with the regular heap have already been covered above (the UPC runtime must ensure that it does not tread into the shared memory area, and if the allocator for the local heap is not reentrant and pthreads are being used, the UPC runtime's wrapper versions of malloc, etc., must perform the needed locking).

The following diagram illustrates the mechanisms used for handling shared memory allocations (for clarity, the diagram is of a UPCR_PURE_DISTRIBUTED configuration, with one process per node: for a threaded implementation there would be two shared regions per process, as in the layout diagram above):

Distributed memory heaps

Each UPC thread has its own shared memory area, containing a local heap growing upward from the bottom, and a global heap growing down from the top. The size of the shared memory area for all threads will be the same, and will be allocated at startup.

Each thread maintains a separate allocator for its local shared heap, along with a handle-safe lock to protect it from concurrent accesses (since memory allocated via upc_alloc may have upc_free called on it asynchronously from another node, locking is required even for local shared heaps).

The global shared memory heaps of all threads are managed by a single allocator on thread 0, which makes decisions on behalf of all the threads. The layout of used/unused areas of the global portion is consistent across all the threads; although the starting/ending virtual addresses of the global heaps may be different (on pthreads implementations it must be, for instance) the size of the heap, and the offset of allocations within it will always the same for all global heaps in a UPC job.

Thread 0 also maintains two counters ('global_size' and 'local_size' in the diagram) that track the number of pages that have been handed out for the global and local heaps, respectively (for the local heap size, this is the size of the largest local heap of all the threads in the UPC job). These counters are monotonically increasing, since pages are never taken away from the heap sizes, only added. The counters exist to make sure that the global and local heaps never collide (if they are about to collide this is a fatal runtime error). The locks are protected by a single lock ("Limit lock" in the diagram): the values may be read without the lock (since they are guaranteed to only increase), but cannot be increased without holding the lock. Note that if both the limit-lock and either the global or local allocator locks on thread 0 need to be held at the same time, the limit lock must be grabbed after the local or global lock: this simple lock hierarchy prevents deadlock (we do not anticipate ever needing to hold all three locks at once, but if we do, then we'll need a lock hierarchy between the local and global locks as well).

Threads other than thread 0 will keep a local copy of the value of local_size ('cached_lsize' in the diagram). This value need not be in sync with the actual value of local_size, but provides a reliable minimum local stack size that local threads can rely on as valid, eliminating the need for some network traffic. This value does not need a lock (it can only be modified by upcr_local_allocs, which are guaranteed to be performed serially by any one thread).

With this diagram in mind, we can now outline pseudocode for each of the UPC runtime memory allocation functions.

 upcr_shared_ptr_t* upcr_local_alloc(int nblocks, int blocksz)
  1. The shared lock for the local allocator is grabbed (this will always succeed: any GASNet asynchronous signal handler that grabs the lock for a upcr_free will complete and release the lock before mainline code can call upcr_local_alloc).
  2. upcra_alloc_sharedlocal(nblocks * blocksz) is called on the local allocator. If it succeeds, the lock is released, and the address received is returned as the function exits.
  3. If upcra_alloc_sharedlocal returns NULL, the allocator is out of memory and needs more pages. The size of the local heap is compared with the cached_lsize: if cached_lsize is greater, the local heap is provided with a number of pages equal to their difference, and the request is tried again.
  4. If the local heap is already as big as cached_lsize, the lock for the allocator is released, and an RPC is made to thread 0 asking for more local stack space. The RPC handler on thread 0 grabs the limit_lock, adds the requested amount to local_size (if needed: the actual value of local_size may already be large enough from other nodes' allocations), and checks to see if local_size + global_size is now larger than the shared memory area size. The lock is then released. If the heaps now overlap, a fatal error is triggered. Otherwise, the RPC returns the new value of local_size to the initiating node, which grabs its local shared allocator lock again, sets its copy of cached_lsize to the returned value, calls upcra_provide_pages_sharedlocal() with the amount of address space it requested, and then calls upcra_alloc_sharedlocal() again (which is guaranteed to be successful this time: only asynchronous upcr_free() calls can have occurred during the RPC), releases the lock, and returns the allocated memory address.

    So, for instance, in the diagram above, if Node 1 needs more pages for its local allocator, it may be able to give them to it without any network traffic, since it has a cached_lsize that exceeds its heap size. Node 2, however, will need to talk to node 0 to increase its local heap size, since it is already using all of the local heap space that it knows is valid.

    Possible optimization: if an RPC request for more memory space is required, it may make sense to ask for more memory than is immediately needed, on the principle that allocations tend to come in groups, hence another allocation is likely to occur soon.

 upcr_shared_ptr_t* upcr_all_alloc(int nblocks, int blocksz)
  1. All threads besides thread 0: simply wait for the result of the function call to be broadcast by thread 0. This is all that these threads need to do, so the remaining steps are only taken by thread 0.
  2. Thread 0: if nblocks is equal to 1, the allocation can be performed entirely on thread 0 (and thus is better performed on the shared local stack rather than the global one, to save space). Thus upcr_local_alloc is called, and the address it returns is broadcast to all nodes.
  3. If the allocation cannot be optimized into a local allocation, the lock for the global allocator is grabbed.
  4. upcra_alloc_sharedglobal is called. If the function returns a valid address, the lock is released, the address broadcast to all of the other threads, and the function returns.
  5. If the allocator returns NULL, it needs more pages: the "Limit lock" is grabbed, the global_size increased by the needed amount, and the value of global_size + local_size is checked to see if it is larger than the shared memory size. If so, a fatal error occurs. Otherwise, the limit lock is released, the new pages are fed to the global allocator via upcra_provide_pages_sharedglobal(), the allocation is made, and the resulting address broadcast to all the nodes.

    Possible optimization: each thread could "touch" (write) a location on each page of the newly allocated region before returning, thereby forcing all threads to page fault (if necessary) at the same time, maintaining better lockstep across processors. For very large allocations (where the size is on the same order as the physical memory size) this optimization would be skipped to prevent introducing thrashing behavior.

 upcr_shared_ptr_t* upcr_global_alloc(int nblocks, int blocksz)
  1. If the thread making this call is not thread 0, an RPC is made to thread 0 to perform the request.
  2. Thread 0: the request is handled by thread 0 in exactly the same way as in upcr_all_alloc (it is transformed into a upcr_local_alloc if the requested memory will all reside on thread 0, otherwise the memory is allocated from the global allocator). There are only two differences: 1) the received address is sent only to the the requesting thread, rather than being broadcast to all threads. 2) if the RPC arrives asynchronously, it is possible that thread 0 will be in a critical section of code, in which case the request will be delayed until the critical section is exited.

Note on (the absence of) race conditions in shared memory allocation

In the case of upcr_global_alloc, memory is being allocated on all nodes, but only 2 nodes (thread 0 and the requesting node) are involved in any network traffic, and are thus aware of the allocation. Thread 2 can allocate global shared memory on thread 1 and write/read that memory without thread 1 even knowing its memory has been allocated.

It will also be OK if during a upc_all_alloc, one node receives the broadcast address and immediately refers to the new memory on another node which has not yet even received the broadcast address (this is theoretically possible if the network allows messages to be delivered out of order).

These types of "refer to the memory on a node without the node knowing that is has been allocated" scenarios are OK, since GASNet guarantees that the entire shared memory region handed out to each thread at startup is valid to be referenced at any time after wards. It guarantees this interface even for networks which do not actually support allowing the entire shared region to be directly accessible at once in its entirely (see the GASNet specification for details).


Home
Downloads
Documentation
Bugs
Publications
Demos
Contact
Internal

This page last modified on Thursday, 31-Aug-2017 17:19:14 PDT