New RMA Design

From Mpich
Jump to: navigation, search



Goals

The goals of the RMA infrastructure are as follows:

  1. No usage of O(P) data structures, where 'P' is the number of processes. Data structure sizes should be constant and user-configurable. General rule is that the user should be able to use more memory to achieve better performance.
  2. Reduce redundant messages where possible (e.g., extra acknowledgment packets).
  3. Reduce process synchronization where possible (e.g., FENCE need not always be synchronizing).
  4. Ensure code is thread-safe (e.g., avoid thread-private data structures).
  5. Ensure integration of RMA code into the overall MPICH progress engine.
  6. Use hardware support for RMA.
  7. Resource management to prevent using up all resources.

Data structures

1. Overview:

We use a new 3D data structure, Operation Table, to store posted RMA operations.

Op-list-slots.jpg

There are the following kinds of data structures involved:

(1) Operation ElementCircle.jpg: contains all origin/target information needed for this RMA operation, plus a request handle that is set when the operation is issued (and has not completed immediately). When the user posts a new operation, the runtime creates and enqueues a new operation structure to the corresponding operation list; when this operation is completed, the runtime dequeues and frees the operation structure from the list.

If req is a NULL pointer, this operation is a PENDING operation; if req is a non-NULL pointer, this operation is a ISSUED operation.

Note that currently operation list is a double-direction linked list, the prev pointer of head points to the tail, and the next pointer of tail is NULL. We may need to make op list as a single-direction linked list, and store tail pointer in target element.

 struct OP_ELEM {
   all OP informations;
   MPID_Request *req;
   struct OP_ELEM *next, *prev;
 }

(From Prof. Gropp's comments) We may need to consider lightweight request design for RMA. Current version of request has higher overhead. For example, we should not always initialize/erase all areas when creating/freeing a request, but should only initialize/erase areas we needed.

(2) Target ElementRhombus.jpg: contains pointer to an operation list that stores all RMA operations to the same target. It also contains the epoch state of the target in cases where a per-target epoch is used (details below). When the origin first communicates with a target, MPICH creates and enqueues a new target element to the corresponding target list. When the origin finishes communication with that target, or all internal resources for targets are used up, MPICH dequeues and frees the target element.

 struct TARGET_ELEM {
   struct OP_ELEM *op_list;
   struct OP_ELEM *next_op_to_issue;
   stuct TARGET_ELEM *next;
   int rank;
   int ack_counter;   // for passive target
   int op_count;
   int sync_flag; // UNLOCK/FLUSH/FLUSH_LOCAL, for piggybacking
   int lock_state; // LOCK_CALLED/LOCK_ISSUED/LOCK_GRANTED 
   int lock_type; // SHARED/EXCLUSIVE
   int lock_mode; // MODE_NO_CHECK
 }

(3) SlotSlot.jpg: contains pointer to a target list. Distribution of targets among slots is currently round-robin. During window creation, MPICH allocates a fixed-sized slot array on the window (size of slot array is controllable by the user).

 struct SLOT {
   struct TARGET_ELEM * targets_list;
 }

(4) WindowWin.jpg: contains at least the following information:

 struct WIN {
   int win_state;  // window state
   int mode;  // for MODE_NOCHECK for LOCKALL case
   int ack_counter;   // for active target
   int num_nonempty_slots;   // how many slots have non-NULL target list to avoid traversing slots
 }


2. Performance issues

For every RMA operation routine, MPICH needs to search in the corresponding target list to find the correct target, which may introduce some overhead when posting operations. However, with a fixed size of the slot array, the overhead of looking up is linear with the number of targets the origin is actively communicating with. For applications that only communicate with a small number of processes, the operation posting time is usually constant. But if the application is communicating with a large number of processes or with more than one process that map to the same slot, there is a linear lookup for the target list within the slot, which can cause overhead.

3. Potential benefits

  1. Operations to different targets are not mixed in a single list, but are stored in separate lists. This has many benefits. For example, for single target epochs, we can judge if all operations to a target have completed (e.g., during a FLUSH operation), and we can find the last operation to a target easily without counting the total number of operations. For multiple targets epoch, we can optimize the garbage collection function by stopping the search in the current target's operation list if we meet incomplete operations (since it is likely that the following operations in that list are also incomplete) and jump to another target's operation list.
  2. Scalability: the size of slots array is fixed and set by the user and is scalable when handling large number of processes on the window.
  3. The garbage collection function can easily distinguish between operations that are not issued and operations that are issued but not completed yet.

Operation and Target Element Pools

We use two-level element pools for the operation elements and the target elements: window (local) and global. The local pool resources can only be used by operations within the window, while the global pool resources are shared by all windows. The sizes of both the local and global portions of each pool are fixed and can be configured by the user at runtime.

When a new element is needed, we first check if one is available in the corresponding local pool. If the local pool is empty, we try to find one in the global pool. If even the global pool is empty, we will call the CLEANUP_WIN_AGGRESSIVE function (see Basic routines) to free up existing elements. When we are done using an element, we first return it to the local pool. If the local pool is full, we return it to the global pool.

This model ensures that a window can never be starved because other windows are using up all the resources.

TODO: We should investigate allocating op/target pools on shared memory, explore if we can make progress on other processes.

Epoch States

RMA epoch states are of two forms: window states and per-target states. Window states are states that affect all targets on the window, while per-target states only affect the individual target they are set for.

There are nine window states:

  1. MPIDI_RMA_NONE
  2. MPIDI_RMA_FENCE_ISSUED
  3. MPIDI_RMA_FENCE_GRANTED
  4. MPIDI_RMA_PSCW_ISSUED
  5. MPIDI_RMA_PSCW_GRANTED
  6. MPIDI_RMA_LOCK_ALL_CALLED
  7. MPIDI_RMA_LOCK_ALL_ISSUED
  8. MPIDI_RMA_LOCK_ALL_GRANTED
  9. MPIDI_RMA_PER_TARGET

There are three per-target states:

  1. MPIDI_RMA_LOCK_CALLED
  2. MPIDI_RMA_LOCK_ISSUED
  3. MPIDI_RMA_LOCK_GRANTED

Note that per-target state is used only when window state is either MPIDI_RMA_LOCK_ALL_CALLED or MPIDI_RMA_PER_TARGET.

Basic routines

1. ISSUE_OPS_TARGET / ISSUE_OPS_WIN / ISSUE_OPS_GLOBAL: Nonblocking call. They try to issue all pending operations for one target/one window/all windows as many as possible. They return number of operations issued.

Note that for ISSUE_OPS_WIN, the function will issue operations in a round-robin fashion to prevent overloading one target with lots of messages at once.

2. CLEANUP_TARGET / CLEANUP_WIN / CLEANUP_GLOBAL: Nonblocking call. They try to find completed operations and targets and clean them up for one target/one window/all windows as many as possible. They return number of operations + targets cleaned up. For active target, this function also cleans up empty target elements. Note that for passive target, empty target functions are not cleaned up by this function and need to be cleaned up by the corresponding packet handler. This is because in passive target, the user might issue a per-target FLUSH operation as well, in which case we need to know how many flush acknowledgments we are waiting for from a particular target.

3. MAKE_RMA_PROGRESS_TARGET / MAKE_RMA_PROGRESS_WIN / MAKE_RMA_PROGRESS_GLOBAL: Nonblocking call. They try to issue pending operations and clean up issued operations for one target/one window/all windows as many as possible.

 MAKE_RMA_PROGRESS_GLOBAL () {
   do {
     x = ISSUE_OPS_GLOBAL();
     y = CLEANUP_OPS_GLOBAL();
   } while (x+y != 0);
 }

MAKE_RMA_PROGRESS_GLOBAL is called from progress engine; MAKE_RMA_PROGRESS_WIN is called in multiple target cases (FENCE/PSCW/LOCK_ALL/FLUSH_ALL); MAKE_RMA_TARGET is called in single target cases (LOCK/FLUSH).

4. PROGRESS_WAIT: it is a blocking call. It keeps poking progress engine until a completion signal is caught in the progress engine. Note that for multithreading case, one thread which is blocked in the function may yield CPU to other threads.

5. CLEANUP_WIN_AGGRESSIVE: it is a blocking call. It tries to make progress and poke progress engine until there are resources available.

 CLEANUP_WIN_AGGRESSIVE(resource_type) {
   if (window_state == FENCE_ISSUED) {
     do PROGRESS_WAIT until the resource is available or window_state == FENCE_GRANTED;
     if (resource is available)
         return;
   }
   
   if (window_state == PSCW_ISSUED) {
     do PROGRESS_WAIT until the resource is available or window_state == PSCW_GRANTED;
     if (resource is available)
         return;
   }
   
   if (window_state == LOCK_ALL_ISSUED) {
     do PROGRESS_WAIT until the resource is available or window_state == LOCK_ALL_GRANTED;
     if (resource is available)
         return;
   }
   
   if (window state == FENCE_GRANTED || window state == PSCW_GRANTED) {
     pick one Target Element K;   // might differ based on resource type
     K->sync_flag = FLUSH_LOCAL;
     call PROGRESS_WAIT until the resource is available;
     return;
   }
   
   if (window state == PER_TARGET || window_state == LOCK_ALL_CALLED || window_state == LOCK_ALL_GRANTED) {
     pick one Target Element K in priority order LOCK_GRANTED, LOCK_ISSUED, LOCK_CALLED;  // might differ based on resource type
     if (resource_type == op element)
         K->sync_flag = FLUSH_LOCAL;
     else   // target element is only freed in the packet handler for passive target
         K->sync_flag = FLUSH;
     call PROGRESS_WAIT until the resource is available;
     return;
   }
 }

Note that for multiple target cases (FENCE/PSCW) we only do a FLUSH_LOCAL because we can track the total number of RMA_DONE packets in the window (win_ptr->ack_counter) and wait for all of them in closing synchronization calls. However, we cannot do the same thing for passive target epochs because we need to know how many RMA_DONE_ACK packets we are expecting from a particular target if the user calls a FLUSH operation.


Algorithm for FENCE

Fence-states.jpg Legends.jpg

Note that there is a request pointer on window called fence_req which stores request of IBARRIER. For details on the memory barrier semantics, see RMA + shared memory.

 FENCE() {
   // If a previous request exists, clean it up
   if (win_ptr->fence_req != NULL) {  // A fence has been previously issued
     if (MPI_MODE_NOPRECEDE) {
       decrement ref count of fence_req by 1;  // let the progress engine delete this request
       set win_ptr->fence_req to NULL;
     }
     else {
       do PROGRESS_WAIT until previous IBARRIER has completed;  // fence_req on window is deleted now
       free win_ptr->fence_req and set it to NULL;
       set window state to FENCE_GRANTED;
     }
   }
   
   // Synchronize as needed
   if (MPI_MODE_NOPRECEDE) { // No operations to complete
     if (SHM is allocated) perform a local BARRIER;  // So that we can issue local operations immediately
     do a global IBARRIER, store the request to fence_req on window;
     set window state to FENCE_ISSUED;
   }
   else { // some operations need to be completed
     do MAKE_RMA_PROGRESS to issue all operations in Operation Table;
     issue (or piggyback) RMA_DONE packets to all active targets and increment ack_counter;
     do PROGRESS_WAIT until Operation Table is empty; // all operations are completed;
     do PROGRESS_WAIT until I received RMA_DONE_ACK packets from all active targets (decrement ack_counter, wait until it reaches 0);
     do a global IBARRER;
     set window state to FENCE_ISSUED; // I am done with all my outgoing messages
     do PROGRESS_WAIT until IBARRIER is completed;  // fence_req on window is freed now
     set window state to FENCE_GRANTED; // I know that everyone else is done
   }
   if (MPI_MODE_NOSUCCEED) { // No operations will start after this
     set window state to UNSET;
   }
 }
 RMA_OP() {
   // Issue local operations immediately
   if (SHM is allocated && target is on the same node) {
       issue operation directly
       return;
   }
   
   // Non-local operation
   do {
       if (window state is FENCE_GRANTED) {
           if (#ops <= PIGGYBACK_THRESHOLD)
               queue up operation and break;  // might need to wait for target and op elements to be available
           else
               issue operation and break;  // if issue didn't complete, might need to queue up request
       }
       else if (window state is FENCE_ISSUED) {
           do progress wait till IBARRIER has completed || (we have a target element && a op element);
           if (IBARRIER completed) {
               set window state to FENCE_GRANTED
               continue;
           }
           else {
               queue up operation and break;  // might need to wait for target and op elements to be available
           }
       }
    } while (1);
 }

Note that if the very last FENCE is not called with MPI_MODE_NOSUCCEED, the window state is still FENCE_GRANTED, other RMA synchronizations will directly transit the window state to correct states; if the very last FENCE is called with MPI_MODE_NOPRECEDE, the window state is still FENCE_ISSUED, other RMA synchronizations will also directly transit the window state to correct states, however, fence_req on window is not freed yet, which will be freed either when we meet another MPI_WIN_FENCE or when we reach MPI_WIN_FREE.


Algorithm for POST-START-COMPLETE-WAIT

Pscw-states.jpg Legends.jpg

The algorithm for PSCW is roughly the same with the algorithm for FENCE, except for the following points:

(1) Every place we did a IBARRIER in the FENCE algorithm, we now do ISENDs (on the target) and IRECVs (on the origin) among the group of processes calling PSCW.

(2) In WIN_POST, if SHM is allocated, the targets wait until all local ISENDs have completed; similarly, in WIN_START, if SHM is allocated, the origins wait until all local IRECVs have completed.

(3) Every place we waited for the IBARRIER request to complete in the FENCE algorithm, we now wait for an array of requests (from ISENDs/IRECVs) to complete instead. The request array is allocated on the window. The size of this array is as large as the process group size used in PSCW, and can be large if the application uses PSCW with a large number of processes. However, such applications should be using FENCE instead of PSCW.

(From Prof. Gropp's comments) For middle scale applications, it is not good to always do direct send/recv for synchronization. We should consider algorithms (like tree-based algorithm) to optimize such cases.

Note that all states are only for origin processes, for target processes, the state is always MPIDI_RMA_NONE.

Algorithm for LOCK-UNLOCK

In this algorithm, a target element might be created in two situations: (1) in the WIN_LOCK call the target element might be created for the first time (it is created in the LOCK_CALLED stated), or (2) in an RMA operation or WIN_UNLOCK call the target element might need to be recreated because it was cleaned up when someone else needed the resource (it is created in the LOCK_GRANTED state).

When the target element is recreated in the second case above, we lose the lock information such as the type of lock (SHARED/EXCLUSIVE) and mode (e.g., MODE_NOCHECK) for this target. The lock type (SHARED/EXCLUSIVE) information is not necessary because we do not need it once we are in the LOCK_GRANTED state. However the lack of the lock mode (MODE_NOCHECK) can hurt performance. This is because, when MODE_NOCHECK is set, we do not acquire the lock and directly set the epoch state to LOCK_GRANTED. However, since we lost this information in the recreation, we might end up sending an UNLOCK packet even when the target was not explicitly locked. In this case, the target is expected to process the UNLOCK message and send back an acknowledgment to the origin.

Note that we free the target element only after we sent the RMA_DONE message and received the RMA_DONE_ACK message from that target.

Lock-states.jpg Legends.jpg

WIN_LOCK() {
   set window state to PER_TARGET;
   if (target_rank == my_rank) {
     do PROGRESS_WAIT until I got the lock;
     return;
   }
   if (SHM is allocated && target is a local process) {
     send lock request message;
     do PROGRESS_WAIT until receiving lock granted message;
     return;
   }
   create Target Element and queue it up;  // might need to restore target elements if none are available
   if (MPI_MODE_NOCHECK)
     set per-target state to LOCK_GRANTED;
   else
     set per-target state to LOCK_CALLED;
 }
 RMA_OP() {
   // Issue local operations immediately
   if (SHM is allocated && target is on the same node) {
       issue operation directly
       return;
   }
   
   // Non-local operation
   do {
       if (target queue does not exist || target state is LOCK_GRANTED) {
           if (#ops <= PIGGYBACK_THRESHOLD)
               queue up operation and break;  // might need to wait for target and op elements to be available
           else
               issue operation and break;        // if issue didn't complete, might need to queue up request
       }
       else if (target state is LOCK_ISSUED) {
           do progress wait till state becomes LOCK_GRANTED || (we have a target element && a op element);
           if (state is LOCK_GRANTED) {
               continue;
           }
           else {
               queue up operation and break;  // might need to wait for target and op elements to be available
           }
       }
       else if (target state is LOCK_CALLED) {
           if (op list is empty && basic datatype && size <= single_op_threshold)
               queue up operation and break;  // might need to wait for target and op elements to be available
           else {
               issue lock operation;
               set state to LOCK_ISSUED;
               continue;
           }
       }
    } while (1);
 }
 WIN_UNLOCK() {
   if (target element for this target does not exist)
       create Target Element and queue it up;  // might need to restore target elements if none are available
   
   if (window state is LOCK_CALLED) {  // single operation optimization
     issue LOCK-OP-UNLOCK to this target;
     do PROGRESS_WAIT until RMA_DONE_ACK is received;
     free this Target Element;
     set window state to UNSET;
     return;
   }
   if (window state is LOCK_ISSUED) {
     do PROGRESS_WAIT until lock granted message is received;
     set per-target state to LOCK_GRANTED;
   }
   if (window state is LOCK_GRANTED) {
     do MAKE_RMA_PROGRESS to issue all operations in Operation Table;
     issue (or piggyback) RMA_DONE+UNLOCK packet to target;
     do PROGRESS_WAIT until operation list for this target is empty; // all operations are completed;
     do PROGRESS_WAIT until I received RMA_DONE_ACK packet from target;
     free this Target Element;
     set window state to UNSET;
   }
 }


Algorithm for LOCK_ALL-UNLOCK_ALL

When the application issues a WIN_LOCK_ALL call, we set the window state to LOCK_ALL_CALLED. No additional messages are issued at this point.

We use two different protocols for LOCK_ALL-UNLOCK_ALL:

(1) Per-target protocol: When the application issues an RMA op to a target, we lazily issue a lock message to that target and set that target's state to LOCK_ISSUED. This protocol has the advantage that no unnecessary lock messages are issued unless the origin talks to a target process. However, this protocol only works till we have sufficient target element resources available. Once we are out of target element resources, we cannot free an existing target element resource because once freed, if we see another RMA operation to that target, we cannot distinguish whether we issued the lock message to that target or not.

(2) Window protocol: When the application runs out of target element resources in the per-target protocol, we fall back to the window protocol, where the window state is changed to LOCK_ALL_ISSUED, a lock operation is issued to all targets to whom a lock operation has not been issued yet, and we wait till the lock acknowledgments for all of these targets arrives. After this the window state is set to LOCK_ALL_GRANTED, at which point we hold a lock to all targets. In this case, target element resources can be freed as needed. While the window protocol adds more lock messages and more synchronization with processes, it is more scalable with respect to resources.

Note that switching from the per-target protocol to the window protocol needs to be handled without requiring the allocation of additional operation or target elements. This can be handled by issuing a constant number of lock operations and keeping track of the associated requests in the function stack (in whichever function triggered the state change from LOCK_ALL_CALLED to LOCK_ALL_ISSUED). Once these requests are free, we issue the next set of requests till all locks are issued and granted.

Also note that when WIN_LOCK_ALL with MODE_NO_CHECK is called, we directly go into the window protocol (but do not need to wait for lock granted messages). The window state is set to LOCK_ALL_GRANTED immediately.

Lock-all-states.jpg Legends.jpg

 LOCK_ALL() {
   if (MPI_MODE_NO_CHECK)
     set window state to LOCK_ALL_GRANTED;
   else
     set window state to LOCK_ALL_CALLED;
 }
 RMA_OPS() {
   if (window state is LOCK_ALL_CALLED) {  // per-target protocol
       if (target queue is available) {
           // follow LOCK/UNLOCK protocol
           MPI_WIN_LOCK(target);
           call RMA_OP on that target;
           return;
       }
       else {
           // fallback to window protocol
           change state to LOCK_ALL_ISSUED
           Issue lock to all targets to which it was not issued
           Wait for all lock acks to come back
           change state to LOCK_ALL_GRANTED
           if (#ops <= PIGGYBACK_THRESHOLD)
              queue up operation and break;  // might need to wait for target and op elements to be available
           else
              issue operation and break;        // if issue didn't complete, might need to queue up request
       }
   }
 }
 UNLOCK_ALL () {
   if (window state is LOCK_ALL_CALLED) {  // per-target protocol
       call MAKE_RMA_PROGRESS to issue all operations in active Target Elements; // for Targets that lock is not issued/granted, wait for lock to be granted;
       issue (or piggyback) RMA_DONE+UNLOCK to active targets;
       call PROGRESS_WAIT until all operations are completed;
       call PROGRESS_WAIT until RMA_DONE_ACK messages from active targets are arrived;
       free all Target Elements;
       set window state to UNSET;
   }
 
   if (window state is LOCK_ALL_GRANTED) {  // window protocol
     call MAKE_RMA_PROGRESS to issue all operations in Operation Table;
     issue (or piggyback) RMA_DONE+UNLOCK to every process on window;
     call PROGRESS_WAIT until all operations are completed;
     call PROGRESS_WAIT until RMA_DONE_ACK messages from all processes are arrived;
     free all Target Elements;
     set window state to UNSET;
   }
 }


Algorithm for FLUSH

WIN_FLUSH is roughly the same with WIN_UNLOCK, except that it doesn't issue (or piggyback) UNLOCK message, but only issues (or piggyback) RMA_DONE message. Also it does not change any window state.

When WIN_FLUSH is used with the window state PER_TARGET, if there is no target element, it does not send an RMA_DONE packet because all operations for this target are guaranteed to be completed when the target element was freed.

When WIN_FLUSH is used with window state LOCK_ALL_CALLED / LOCK_ALL_GRANTED, if there is no target element, it will send an RMA_DONE packet and wait for an RMA_DONE_ACK packet, because RMA operations are not guaranteed to have completed on the target.


Performance optimizations

  1. Single short operation optimization: if there is only one operation between the WIN_LOCK and WIN_UNLOCK with a basic datatype and its size is smaller than single_op_opt_threshold, we piggyback both the LOCK and UNLOCK messages on this operation. Specifically, we will not issue the lock request packet or the unlock packet. Instead, we only send one packet at WIN_UNLOCK time which contains lock request, operation data and the unlock flag. We need to wait for the acknowledgment to come back before returning from WIN_UNLOCK.
  2. Piggyback RMA_DONE (UNLOCK) packet: If there are GET operations in the operation list, we move them to the tail and piggyback the RMA_DONE (or UNLOCK) flag on this message. In such cases, the data returned by the GET operation also piggybacks the RMA_DONE_ACK (or UNLOCK_ACK) flag. If there is no GET operations in the operation list, we piggyback the RMA_DONE packet with the last operation and wait for a separate RMA_DONE_ACK packet to arrive. Specifically, we always store the tail of the operation list before reaching the ending synchronization and update it if necessary when we encounter a new RMA operation. When the tail is already a GET/GET_ACCUM/CAS/FOP operation, we do not need to update it anymore. When the tail is a PUT/ACC operation, and if we encounter a GET operation, we update the tail with that GET operation and do not need to update it in future; if we encounter a non-GET operation, we also update the tail with that operation because we need to guarantee the ordering of operations. (If the user application doesn't care about the RMA_DONE packet, they should be able to provide MPI info hints to tell runtime that not always waiting for RMA_DONE packet to come back.)
  3. Lazy deallocation of target elements: In passive target communication, the target structure holds additional information such as the mode of the lock (e.g., MPI_MODE_NOCHECK). This information allows us to decide whether to send an additional UNLOCK message or not. However, if the target element is freed, this information is lost and we might need to fallback to a conservative model of sending UNLOCK packets even when they are not needed. The lazy deallocation of target elements optimization allows us to keep these elements allocated even when all of the associated operations in the queue have been issued and freed. The target element is only freed when we run out of resources or we close that epoch.

Note that the piggyback optimization has the risk of deadlock, when every process queues up operations and used up all operation resources. However, using operation local pool can avoid such situation. When we use up all operation resources, we will call MAKE_RMA_PROGRESS_TARGET to issue out operations, and then call PROGRESS_WAIT until they are finished and we can get operation resources from the local pool. The free function of operation always first puts the operation to the local pool, when local pool is full, it will put the operation to the global pool.

A key shortcoming of all of these performance optimizations is that they hold up resources that would typically only be freed by user intervention. That is, without these optimizations, if a process waited in the progress engine, it can reclaim resources within a finite amount of time (assuming other processes in the system are making MPI calls). However, with these optimizations, this might not be true. For example, if the system has operation elements queued up waiting for an UNLOCK packet (in order to do the single short operation optimization mentioned above), the operation element will not be freed till the user issues the UNLOCK operation (or a flush operation). To handle such situations, we need to distinguish cases of resource exhaustion in the implementation and temporarily disable appropriate optimizations to reclaim resources.

RMA + threads

PROGRESS_WAIT can only happen in the main function calls associated with RMA operations and RMA synchronizations, and we need to make sure that those RMA calls work correctly with multithreading. Interleaving situations that need to be considered are as follows:

1. PUT/GET/ACC + PUT/GET/ACC

2. PUT/GET/ACC + WIN_FENCE (no MODE_NOSUCCEED and no MODE_NOPRECEDE)

3. PUT/GET/ACC + WIN_FLUSH

4. WIN_FLUSH + WIN_FLUSH

All other interleaving situations are invalid.

TODO: needs more careful and detailed definition and clarification on threads interleaving.

Enabling multithreading in WIN_FLUSH: To achieve this, we piggyback the FLUSH_DONE (i.e. RMA_DONE) with the last packet. Specifically, each thread (suppose T1) creates a response request on the origin and passes its handle to the last packet. The completion of that response request is handed over to the progress engine, similar to other operations. The response request is completed and freed on the origin only when the FLUSH_DONE_ACK (i.e. RMA_DONE_ACK) packet arrives, which indicates the completion of WIN_FLUSH on thread T1. Note that the thread that creates the response request (T1) and the thread that completes the response request (suppose T2) are not necessarily the same. As long as the response request created by T1 gets completed, WIN_FLUSH on T1 can return.

RMA + shared memory

When SHM is allocated for RMA window, we need to add memory berriers at proper places in RMA synchronization routines to guarantee the ordering of read/write operations, so that any operations after synchronization calls will see the correct data.

There are four kinds of operations involved in the following explanation:

(1) Local loads/stores: any operations happening outside RMA epoch and accessing each process's own window memory.

(2) SHM operations: any operations happening inside RMA epoch. They may access any processes' window memory, which include direct loads/stores, and RMA operations that are internally implemented as direct loads/stores in MPI implementation.

(3) PROC_SYNC: synchronizations among processes by sending/receiving messages.

(4) MEM_SYNC: a full memory barrier. It ensures the ordering of read/write operations on each process.


Followings are explanations about when memory barriers should be called in different RMA synchronizations.

1. FENCE synchronization

              RANK 0                           RANK 1
     
      (local loads/stores)            (local loads/stores)
     
          WIN_FENCE {                      WIN_FENCE {
              MEM_SYNC                       MEM_SYNC
              PROC_SYNC -------------------- PROC_SYNC
              MEM_SYNC                       MEM_SYNC
          }                              }
   
       (SHM operations)                (SHM operations)
   
          WIN_FENCE {                    WIN_FENCE {
              MEM_SYNC                       MEM_SYNC
              PROC_SYNC -------------------- PROC_SYNC
              MEM_SYNC                       MEM_SYNC
          }                              }
   
     (local loads/stores)             (local loads/stores)

We need MEM_SYNC before and after PROC_SYNC for both starting WIN_FENCE and ending WIN_FENCE, to ensure the ordering between local loads/stores and PROC_SYNC in starting WIN_FENCE (and vice versa in ending WIN_FENCE), and the ordering between PROC_SYNC and SHM operations in starting WIN_FENCE (and vice versa for ending WIN_FENCE).

In starting WIN_FENCE, the MEM_SYNC before PROC_SYNC essentially exposes previous local loads/stores to other processes; after PROC_SYNC, each process knows that everyone else already exposed their local loads/stores; the MEM_SYNC after PROC_SYNC ensures that my following SHM operations will happen after PROC_SYNC and will see the latest data on other processes.

In ending WIN_FENCE, the MEM_SYNC before PROC_SYNC essentially exposes previous SHM operations to other processes; after PROC_SYNC, each process knows everyone else already exposed their SHM operations; the MEM_SYNC after PROC_SYNC ensures that my following local loads/stores will happen after PROC_SYNC and will see the latest data in my memory region.

2. POST-START-COMPLETE-WAIT synchronization

             RANK 0                        RANK 1
  
                                    (local loads/stores)
          
          WIN_START {                  WIN_POST {
                                           MEM_SYNC
              PROC_SYNC ------------------ PROC_SYNC
              MEM_SYNC
          }                             }
          
        (SHM operations)
  
          WIN_COMPLETE {              WIN_WAIT/TEST {
              MEM_SYNC
              PROC_SYNC ----------------- PROC_SYNC
                                          MEM_SYNC
          }                             }
    
                                    (local loads/stores)

We need MEM_SYNC before PROC_SYNC for WIN_POST and WIN_COMPLETE, and MEM_SYNC after PROC_SYNC in WIN_START and WIN_WAIT/TEST, to ensure the ordering between local loads/stores and PROC_SYNC in WIN_POST (and vice versa in WIN_WAIT/TEST), and the ordering between PROC_SYNC and SHM operations in WIN_START (and vice versa in WIN_COMPLETE).

In WIN_POST, the MEM_SYNC before PROC_SYNC essentially exposes previous local loads/stores to group of origin processes; after PROC_SYNC, origin processes knows all target processes already exposed their local loads/stores; in WIN_START, the MEM_SYNC after PROC_SYNC ensures that following SHM operations will happen after PROC_SYNC and will see the latest data on target processes.

In WIN_COMPLETE, the MEM_SYNC before PROC_SYNC essentailly exposes previous SHM operations to group of target processes; after PROC_SYNC, target processes knows all origin process already exposed their SHM operations; in WIN_WAIT/TEST, the MEM_SYNC after PROC_SYNC ensures that following local loads/stores will happen after PROC_SYNC and will see the latest data in my memory region.

3. Passive target synchronization

Note that for passive target synchronization, when one process wants to access its own window memory, it must do those accesses within an passive epoch (make local loads/stores become SHM operations), or using MPI_WIN_SYNC before and after local loads/stores, otherwise those local accesses may be reordered and leads to wrong results (clarification is needed for this in MPI specification). See following for details.

             RANK 0                          RANK 1
   
                                       WIN_LOCK(target=1) {
                                           PROC_SYNC (lock granted)
                                           MEM_SYNC
                                       }
  
                                       (SHM operations)
  
                                       WIN_UNLOCK(target=1) {
                                           MEM_SYNC
                                           PROC_SYNC (lock released)
                                       }
  
        PROC_SYNC -------------------- PROC_SYNC
  
        WIN_LOCK (target=1) {
            PROC_SYNC (lock granted)
            MEM_SYNC
       }
 
        (SHM operations)
 
        WIN_UNLOCK (target=1) {
            MEM_SYNC
            PROC_SYNC (lock released)
        }
 
        PROC_SYNC -------------------- PROC_SYNC
 
                                       WIN_LOCK(target=1) {
                                           PROC_SYNC (lock granted)
                                           MEM_SYNC
                                       }
  
                                       (SHM operations)
  
                                       WIN_UNLOCK(target=1) {
                                           MEM_SYNC
                                           PROC_SYNC (lock released)
                                       }
  

We need MEM_SYNC after PROC_SYNC in WIN_LOCK, and MEM_SYNC before PROC_SYNC in WIN_UNLOCK, to ensure the ordering between SHM operations and PROC_SYNC and vice versa.

In WIN_LOCK, the MEM_SYNC after PROC_SYNC guarantees two things: (a) it guarantees that following SHM operations will happen after lock is granted; (b) it guarantees that following SHM operations will happen after any PROC_SYNC with target before WIN_LOCK is called, which means those SHM operations will see the latest data on target process.

In WIN_UNLOCK, the MEM_SYNC before PROC_SYNC also guarantees two things: (a) it guarantees that SHM operations will happen before lock is released; (b) it guarantees that SHM operations will happen before any PROC_SYNC with target after WIN_UNLOCK is returned, which means following SHM operations on that target will see the latest data.

WIN_LOCK_ALL/UNLOCK_ALL are same with WIN_LOCK/UNLOCK.

             RANK 0                          RANK 1
 
        WIN_LOCK_ALL
 
        (SHM operations)
 
        WIN_FLUSH(target=1) {
            MEM_SYNC
        }
 
        PROC_SYNC ----------------------- PROC_SYNC
 
                                          WIN_LOCK(target=1) {
                                              PROC_SYNC (lock granted)
                                              MEM_SYNC
                                          }
 
                                          (SHM operations)
 
                                          WIN_UNLOCK(target=1) {
                                              MEM_SYNC
                                              PROC_SYNC (lock released)
                                          }
 
        WIN_UNLOCK_ALL

We need MEM_SYNC in WIN_FLUSH to ensure the ordering between SHM operations and PROC_SYNC.

The MEM_SYNC in WIN_FLUSH guarantees that all SHM operations before this WIN_FLUSH will happen before any PROC_SYNC with target after this WIN_FLUSH, which means SHM operations on target process after PROC_SYNC with origin will see the latest data.

User can also use MPI_WIN_SYNC in passive target synchronization to guarantee the ordering of load/store operations.

             RANK 0                          RANK 1
   
                                         WIN_SYNC {
                                             MEM_SYNC
                                         }
  
                                       (local loads/stores)
  
                                         WIN_SYNC {
                                             MEM_SYNC
                                         }
  
        PROC_SYNC -------------------- PROC_SYNC
  
        WIN_LOCK (target=1) {
            PROC_SYNC (lock granted)
            MEM_SYNC
       }
 
        (SHM operations)
 
        WIN_UNLOCK (target=1) {
            MEM_SYNC
            PROC_SYNC (lock released)
        }
 
        PROC_SYNC -------------------- PROC_SYNC
 
                                       WIN_SYNC {
                                           MEM_SYNC
                                       }
  
                                       (local loads/stores)
  
                                       WIN_SYNC {
                                           MEM_SYNC
                                       }

Reduce O(p) structure on window

This includes sharing window information within the node and local cache for remote window information.

(From Prof. Gropp's comments)

  1. Detect common arguments (e.g., the same displacement unit or size used on all processes). Possible enhancement: consider as an integer map, and share the group compression code. Ditto other data.
  2. Based on CH3 channel needs, determine whether offsets are needed for all process or just the local process. In the case where no direct RMA is available, there is no need for all processes to hold

the offsets.

  1. For data that must be available, consider caching and/or accessing with one-sided operations (e.g., only hold the window base addresses for the windows that you are accessing; have a way to handle a cache "miss").
  2. Related to 3, store shared information in a single, shared data structure for all processes on the same SMP. Consider extending this approach to other MPI objects.

Datatype in RMA

(From Prof. Gropp's comments)

The current RMA code doesn’t interface correctly to the dataloop/datatype code. The problem is that it apparently expects a “flattened” version of the dataloop, which can be terrible for performance. At the least, we should consider the following:

  1. Contiguous must be very fast
  2. Strided should be fast (e.g., vector types); this probably means a special case for this
  3. Datatypes used at the target may be cached and do not need to be sent every time
    1. Caches can be shared on nodes.
    2. Caches need to flush data, so it may be necessary to refresh a cached item
  4. Non contiguous data should be pipelined if large enough, rather than packed first and then sent in a contiguous lump.
  5. This must be integrated with one-sided support in the interconnect hardware.

Ordering

MPICH provides a netmod API (get_ordering) to allow netmod to expose the ordering of Active-Messages-based (AM-based) operations (am_ordering). A netmod may issue some packets via multiple connections in parallel (such as RMA), and those packets can be unordered. In such case, the netmod should return 0 from get_ordering function; otherwise, the netmod returns 1 from get_ordering function. The default value of am_ordering in MPICH is 0. Setting it to 1 may improve the performance of runtime because it does not need to maintain ordering among AM-based operations. One example of implementing the API in MXM netmod is as follows:

  int MPID_nem_mxm_get_ordering(int *ordering)
  {
    (*ordering) = 1;
    return MPI_SUCCESS;
  }

When am_ordering is 1, the MPICH runtime only issues a REQ_ACK at last when it wants to detect remote completion of all operations; when am_ordering is 0, the runtime needs to issue a REQ_ACK with every issued operations to ensure that any future remote completion detection will be correct.