WikiPrint - from Polar Technologies

Accessing Network Attached Memory (NAM) from Tasks using TAMPI

Table of contents:

Quick Overview

Task-Aware MPI (TAMPI)

Accessing NAM through ParaStationMPI and TAMPI

Heat Benchmark

In this section, we exemplify the use of TAMPI and NAM windows through the Heat benchmark. We use an iterative Gauss-Seidel method to solve the Heat equation, which is a parabolic partial differential equation that describes the distribution of heat in a given region over time. This benchmark simulates the heat diffusion on a 2D matrix of floating-point elements during multiple timesteps. The 2D matrix is logically divided into 2D blocks and may have multiple rows and columns of blocks. The computation of an element at position M[r][c] in the timestep t depends on the value of the top and left elements (M[r-1][c] and M[r][c-1]) computed in the current timestep t, and the right and bottom elements (M[r][c+1] and M[r+1][c]) from the previous timestep t-1. We can extrapolate this logic in the context of blocks so that a block has a dependency on the computation of its adjacent blocks. Notice that the computation of blocks in a diagonal is fully concurrent because there is no dependency between them.

There are three different MPI versions, and all of them distribute the 2D matrix across ranks assigning consecutive rows of blocks to each MPI rank. Note that the matrix is distributed by blocks vertically but not horizontally. Therefore, an MPI rank has two neighboring ranks: one above and another below. The exceptions are the first and last ranks since they have a single neighbor. This distribution requires the neighboring ranks to exchange the external rows (halos) from their boundary blocks in order to compute their local blocks in each timestep.

This benchmark is publicly available in the heat subfolder at the ?https://gitlab.version.fz-juelich.de/DEEP-EST/ompss-2-benchmarks repository. The first version is based on an MPI-only parallelization, while the other two are hybrid MPI+OmpSs-2 leveraging tasks and the TAMPI library. We briefly describe each one below:

Using NAM in Heat benchmark

In this benchmark, we use the NAM memory to save the computed matrix periodically. The idea is to save different states (snapshots) of the matrix during the execution in a persistent NAM memory region. Then, another program could retrieve all the matrix snapshots, process them and produce a GIF animation showing the heat's evolution throughout the execution. Notice that we cannot use regular RAM for that purpose because the matrix could be huge, and we may want to store tens of matrix snapshots. We also want to keep it persistently so that other programs could process the stored data. Moreover, the memory should be easily accessible by the multiple MPI ranks or their tasks in parallel. The NAM memory satisfies all these requirements, and as previously stated, ParaStationMPI allows accessing NAM allocations through standard MPI RMA operations. We only implement the NAM snapshots in the TAMPI variants 02.heat_itampi_ompss2_tasks.bin and 03.heat_tampirma_ompss2_tasks.bin.

The Heat benchmark allocates a single MPI window that holds the whole NAM region, which is used by all ranks (via the MPI_COMM_WORLD communicator) and throughout the execution. Every few timesteps (specified by the user), it saves the whole matrix into a specific NAM subregion. Each timestep that saves a matrix snapshot employs a distinct NAM subregion. These subregions are placed one after the other, consecutively, without overlapping. Thus, the entire NAM region's size is the full matrix size multiplied by the number of times the matrix will be saved. Even so, we allocate the NAM memory region using the Managed Contiguous layout (psnam_structure_managed_contiguous). This means that rank 0 allocates the whole region, but each rank acquires a consecutive memory subset, where it will store its blocks' data for every snapshot. For instance, the NAM allocation will first have the space for storing all snapshots of the blocks from rank 0, followed by the space for all snapshots of blocks from rank 1, and so on. By using that layout, NAM subregions are rank-addressed using the rank it belongs to, simplifying the saving and retrieving of snapshots.

Then, when a timestep requires a snapshot, the application instantiates multiple tasks that save the matrix data into the corresponding NAM subregion. Each MPI rank creates a task for writing (copying) the data of each of its matrix blocks into the NAM subregion. These communication tasks do not have any data dependency between them, so they can write data to the NAM using regular MPI_Put in parallel. Ranks only write to their own subregions, never in other ranks' subregions. Nevertheless, we must call all MPI_Puts inside an MPI RMA access epoch, so there must be a window fence call before all the MPI_Puts, and another one after them to close the epoch, for each of the snapshot timesteps. Here is where we leverage the new function MPI_Win_ifence along with the TAMPI non-blocking support. In this way, we can fully taskify both synchronization and writing of the NAM window, keeping the data-flow model and without closing the parallelism for doing the snapshots (e.g., with a taskwait) . Thanks to the task data dependencies and TAMPI, we cleanly include the snapshots in the application's data-flow execution as any other regular communication task with dependencies.

The following pseudo-code shows how the saving of snapshots works in 02.heat_itampi_ompss2_tasks.bin: 

void solve() {
    int namSnapshotFreq = ...;
    int namSnapshotId = 0;

    for (t = 1; t <= timesteps; ++t) {
        // Computation and communication tasks declaring
        // dependencies on the blocks they process
        gaussSeidelSolver(..all blocks in current rank..);

        if (t % namSnapshotFreq == 0) {
            namSaveMatrix(namSnapshotId, namWindow, ...);
            ++namSnapshotId;
        }
    }
    #pragma oss taskwait
}

The function above is the main procedure that executes the Heat's timesteps applying the Gauss-Seidel method. All MPI ranks run this function concurrently, each one working with their corresponding blocks from the matrix. In each timestep, the gaussSeidelSolver function instantiates all the computation and communication tasks that process the rank's blocks and exchanges the halo rows with the neighboring ranks. These tasks declare the proper input/output dependencies on the blocks they are reading/writing. Every some timesteps, the algorithm calls namSaveMatrix that will issue tasks to perform a snapshot of the data computed after computing that timestep. Notice that namSaveMatrix will have to instantiate tasks with input dependencies on the blocks in order to perform the snapshot at the right moment of the execution, i.e. after the gaussSeidelSolver tasks from that timestep. Notice also that we identify each snapshot with the namSnapshotId, which we use to know where we should store the snapshot data inside the NAM window. This taskwait at the end of the algorithm is the only one in the whole execution; we do not close the parallelism anywhere else. 

void namSaveMatrix(int namSnapshotId, MPI_Win namWindow, ...) {
    // Compute snapshot offset inside NAM region
    int snapshotOffset = namSnapshotId*sizeof(..all blocks in current rank..);

    // Open RMA access epoch to write the NAM window for this timestep
    #pragma oss task in(..all blocks in current rank..) inout(namWindow)
    {
        MPI_Request request;
        MPI_Win_ifence(namWindow, 0, &request);
        TAMPI_Iwait(&request, MPI_STATUS_IGNORE);
    }

    // Write all blocks from the current rank to NAM subregions concurrently
    for (B : all blocks in current rank) {
        #pragma oss task in(..block B..) in(namWindow)
        {
            MPI_Put(/* source data */   ..block B..,
                    /* target rank */   currentRank,
                    /* target offset */ snapshotOffset + B,
                    /* target window */ namWindow);
        }
    }

    // Close RMA access epoch to write the NAM window for this timestep
    #pragma oss task in(..all blocks in current rank..) inout(namWindow)
    {
        MPI_Request request;
        MPI_Win_ifence(namWindow, 0, &request);
        TAMPI_Iwait(&request, MPI_STATUS_IGNORE);
    }
}

The function above is the one called periodically from the primary procedure. It instantiates the tasks that will perform the snapshot of the current rank's blocks into their corresponding NAM memory subregions. The first step is to compute the offset of the current snapshot inside the NAM region using the snapshot identifier. Before writing to the NAM window, the application must ensure that an MPI RMA access epoch has been opened on that window. That is what the first task is doing. After all the blocks have been computed in that timestep and are ready to be read (notice its task dependencies), the first task will run and execute an MPI_Win_ifence to start the window epoch's opening. This MPI function generates an MPI request and serves as parameter of the subsequent call to TAMPI_Iwait, which binds the current task's completion to the finalization of the MPI request. This last call is non-blocking and asynchronous, so the fence operation may not be completed after returning. The task can finish its execution, but it will not complete until the fence operation finishes. Once it finishes, TAMPI will automatically complete the task and make its successor tasks ready. The successors of the fence task are the ones that perform the actual writing (copying) of data into the NAM memory by calling MPI_Put. All blocks can be saved in the NAM memory in parallel by different tasks. The source of the MPI_Put is the block itself (in regular RAM), while the destination is the place where the block should be written inside the NAM region. After all writer tasks finish, it is the turn for the task that closes the MPI RMA access epoch on the NAM window. This one should behave similarly to the one that opened the epoch.

Notice that all tasks declare the proper dependencies on both the matrix blocks and the NAM window to guarantee their correct execution order. Thanks to these data dependencies and the TAMPI non-blocking feature, we can cleanly add the execution of the snapshots into the task graph, being executed asynchronously, and being naturally interleaved with the other computation and communication tasks. Finally, it is worth noting that the blocks are written into the NAM memory in parallel, utilizing the machine's CPU and network resources efficiently.

Requirements

The requirements of this application are shown in the following lists. The main requirements are:

Building & Executing on DEEP

The instructions to build and execute the Heat benchmark with NAM snapshots will appear here soon.

References