View source for New RMA Design

[[Category:Design Documents]]

== Goals ==

The goals of the RMA infrastructure are as follows:

# 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.
# Reduce redundant messages where possible (e.g., extra acknowledgment packets).
# Reduce process synchronization where possible (e.g., FENCE need not always be synchronizing).
# Ensure code is thread-safe (e.g., avoid thread-private data structures).
# Ensure integration of RMA code into the overall MPICH progress engine.


== Data structures ==

'''1. Overview''': 

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

[[File:Op-list-slots.jpg]]

There are three kinds of data structures involved:

'''(1) Operation Element''': 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.

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

'''(2) Target Element''': 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 * ops_list;
    stuct TARGET_ELEM * prev, * next;
    int rank;
    int ops_count;
    int target_state; // LOCK_CALLED/LOCK_ISSUED/LOCK_GRANTED 
    int target_type; // SHARED/EXCLUSIVE
    int target_mode; // MODE_NO_CHECK
  }

'''(3) Slot''': 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;
  }

Note that there is a variable on the window indicating how many slots have non-NULL target list. It helps optimizing ISSUE_OPS and CLEANUP_OPS (see [https://wiki.mpich.org/mpich/index.php/New_RMA_Design#Algorithms_for_RMA_synchronizations Algorithms for RMA synchronizations]), by avoiding traversing all 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'''

# 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.
# 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. 
# 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 element pools for the operation elements and the target elements.  The pools are created at MPI initialization time and are shared among all windows.  The size of each pool is fixed and can be configured by the user at runtime.

When we need a new element, we first check if one is available in the corresponding pool.  If not, we will call the '''RESTORE_OP_ELEM_RESOURCES''' or '''RESTORE_TARGET_ELEM_RESOURCES''' functions (see [https://wiki.mpich.org/mpich/index.php/New_RMA_Design#Basic_routines Basic routines]) to free up existing elements.


== 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''':

# MPIDI_EPOCH_UNSET
# MPIDI_EPOCH_FENCE_ISSUED
# MPIDI_EPOCH_FENCE_GRANTED
# MPIDI_EPOCH_PSCW_ISSUED
# MPIDI_EPOCH_PSCW_GRANTED
# MPIDI_EPOCH_LOCK_ALL_CALLED
# MPIDI_EPOCH_LOCK_ALL_ISSUED
# MPIDI_EPOCH_LOCK_ALL_GRANTED
# MPIDI_EPOCH_PER_TARGET

There are three '''per-target states''':

# MPIDI_EPOCH_LOCK_CALLED
# MPIDI_EPOCH_LOCK_ISSUED
# MPIDI_EPOCH_LOCK_GRANTED


== 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_OPS_TARGET / CLEANUP_OPS_WIN / CLEANUP_OPS_GLOBAL''': Nonblocking call. They try to find completed operations and clean them up for one target/one window/all windows as many as possible. They return number of operations cleaned up.

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. '''RESTORE_TARGET_ELEM_RESOURCES''': it is a blocking call. It tries to make progress and poke progress engine until there are Target Element resources available.

  RESTORE_TARGET_ELEM_RESOURCES() {
    if (window state == FENCE_GRANTED || window state == PSCW_GRANTED || window state == LOCK_ALL_GRANTED || window state == LOCK_ALL_CALLED) {
      pick one Target Element K;
      do MAKE_RMA_PROGRESS and PROGRESS_WAIT until K’s operation list is empty;
      send (or piggyback) RMA_DONE packet to K;   // do not wait for RMA_DONE_ACK to arrive
      win_ptr->ack_count++;
      free Target Element K;
    }
   
    else if (window state == PER_TARGET) {
      pick one Target Element K in priority order LOCK_GRANTED, LOCK_ISSUED, LOCK_CALLED;
      do MAKE_RMA_PROGRESS and PROGRESS_WAIT until K’s operation list is empty;
      send (or piggyback) RMA_DONE packet to K;   // wait for RMA_DONE_ACK to arrive
      free Target Element K;
    }
  }
 
Note that for multiple target cases (FENCE/PSCW/LOCK_ALL) we send RMA_DONE packets but do not need to wait for RMA_DONE_ACK packets in RESTORE_TARGET_ELEM_RESOURCES, because we can just track total number of RMA_DONE packet and wait for all of them in closing synchronization calls, however, we cannot do the same thing for single target cases (LOCK), because we need an O(P) array to track number of RMA_DONE packets to each target. 
 
6. '''RESTORE_OP_ELEM_RESOURCES''': it is a blocking call. It tries to make progress and poke progress engine until there are Operation Element resources available. It has risk of deadlock due to piggyback optimization (see [https://wiki.mpich.org/mpich/index.php/New_RMA_Design#Performance_optimizations Performance optimizations]).

Suppose every process in the epoch queues up operations in order to do piggyback later in the synchronization. At one point, Operation Elements resources are all used up, and everyone is blocked waiting for resources to be available. Since no one makes progress on issuing operations, resources will never be available again, leads to deadlock situation.
To avoid the deadlock, we need to disable piggyback optimization in order to issue operations. We define a global variable piggyback_flag to indicate if piggyback is currently enabled or not. Following pseudo-code shows when to disable piggyback optimization.

  x = GET_OP_ELEM();
  if (x == NULL) {
    // run out of resources
  
    // first try to make progress on all windows
    MAKE_RMA_PROGRESS_GLOBAL();
  
    x = GET_OP_ELEM();
    while (x == NULL) {
      // if still run out of resources, disable piggyback, make progress
      piggyback_flag = 0;
      PROGRESS_WAIT();    // call MAKE_RMA_PROGRESS_GLOBAL() internally
      x = GET_OP_ELEM();
      piggyback_flag = 1;
    }
  }

== Algorithm for FENCE ==

[[File:Fence-states.jpg]] [[File: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 [https://wiki.mpich.org/mpich/index.php/New_RMA_Design#RMA_.2B_shared_memory RMA + shared memory].

  FENCE() {
    // Memory barrier for shared memory operations
    if (SHM is allocated) do memory barrier;
    
    // 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;
    }
    
    // Memory barrier for shared memory operations
    if (SHM is allocated) do memory barrier;
  }

  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 ==

[[File:Pscw-states.jpg]] [[File: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.


== 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.

[[File:Lock-states.jpg]] [[File: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 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.

[[File:Lock-all-states.jpg]] [[File:Legends.jpg]]

(Simplified version of pseudo-code)

  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 (window state is LOCK_ALL_GRANTED) {
      if (no Target Element for this target) {
        // create a Target Element for this target, set per-target state to LOCK_GRANTED
      }
    }
  
    if (per-target state is LOCK_CALLED) {
      // if single op optimization is satisfied, queue it up, return;
      // if single op optimization is not satisfied, queue it up, send lock request, set per-target state to LOCK_ISSUED, return;
    }
  
    if (per-target state is LOCK_ISSUED) {
      // queue it up, return;
    }
  
    if (per-target state is LOCK_GRANTED) {
      // if piggyback_threshold is not reached, queue it up, return;
      // if piggyback_threshold is reached, call MAKE_RMA_PROGRESS
    }
  }

  UNLOCK_ALL () {
    while (Operation Table is not empty) {
      call MAKE_RMA_PROGRESS;  // issue as fast as possible (e.g. do not wait on one target(slot) to finish before starting the next target (slot))
      call PROGRESS_WAIT;
    }
    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 issue (or piggyback) RMA_DONE message. Also it does not change any window state.

When WIN_FLUSH is used with window state is 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 Target Element is 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 RMA_DONE_ACK packet, because RMA operations are not guaranteed to be completed on target.

== Performance optimizations ==

1. '''Piggyback'''

# Single short operation optimization: if there is only one operation between WIN_LOCK and WIN_UNLOCK, data type is basic datatype, and data volume is smaller than single_op_opt_threshold, we piggyback LOCK message and UNLOCK message with this operation. Specifically, we will not issue lock request packet and unlock packet, instead, we only send one packet in WIN_UNLOCK which contains lock request, operation data and unlock flag.
# piggyback RMA_DONE (UNLOCK) packet: if there are small number of operations posted, we piggyback the RMA_DONE packet with the last operation. Specifically, when number of queued operations is smaller than '''piggyback_threshold''', we send the lock request packet but continue to queue operations up until reach the ending synchronization. If there are GET operations in the operation list, we move that GET to the tail, in such case, the completion of that GET operation indicates the completion of all RMA operations on target; if there is no GET operations in the operation list, we piggyback RMA_DONE packet with the last operation and waits for RMA_DONE_ACK packet to arrive.

== RMA + threads ==

PROGRESS_WAIT can only happen in main function calls (RMA operations, RMA synchronizations), and we need to make sure that those RMA calls works 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

Except for those, any other interleaving situations are not correct programs, so runtime should not need to deal with them.

'''Enable 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 origin and pass its handle to the last packet. The completion of that response request is handed over to the progress engine now, same with other operations. The response request is completed and freed on origin only when 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 got 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), 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.