Search in sources :

Example 1 with RecoverPartKey

use of com.tencent.angel.ps.ha.RecoverPartKey in project angel by Tencent.

the class MasterService method psReport.

/**
 * response for parameter server heartbeat
 *
 * @param controller rpc controller of protobuf
 * @param request heartbeat request
 */
@SuppressWarnings("unchecked")
@Override
public PSReportResponse psReport(RpcController controller, PSReportRequest request) throws ServiceException {
    if (LOG.isDebugEnabled()) {
        LOG.debug("receive ps heartbeat request. request=" + request);
    }
    // parse parameter server counters
    List<Pair> params = request.getMetricsList();
    int size = params.size();
    Map<String, String> paramsMap = new HashMap<String, String>();
    for (int i = 0; i < size; i++) {
        paramsMap.put(params.get(i).getKey(), params.get(i).getValue());
    }
    PSAttemptId psAttemptId = ProtobufUtil.convertToId(request.getPsAttemptId());
    PSReportResponse.Builder resBuilder = PSReportResponse.newBuilder();
    if (!context.getParameterServerManager().isAlive(psAttemptId)) {
        // if psAttemptId is not in monitor set, just return a PSCOMMAND_SHUTDOWN command.
        LOG.error("ps attempt " + psAttemptId + " is not in running ps attempt set");
        resBuilder.setPsCommand(PSCommandProto.PSCOMMAND_SHUTDOWN);
    } else {
        resBuilder.setPsCommand(PSCommandProto.PSCOMMAND_OK);
        // refresh last heartbeat timestamp
        context.getParameterServerManager().alive(psAttemptId);
        // send a state update event to the specific PSAttempt
        context.getEventHandler().handle(new PSAttemptStateUpdateEvent(psAttemptId, paramsMap));
        // Check is there save request
        PSMatricesSaveContext subSaveContext = context.getModelSaver().getSaveContext(psAttemptId.getPsId());
        PSMatricesSaveResult subSaveResult = context.getModelSaver().getSaveResult(psAttemptId.getPsId());
        if (subSaveContext != null && subSaveResult != null && (subSaveContext.getRequestId() == subSaveResult.getRequestId()) && (subSaveResult.getState() == SaveState.INIT || subSaveResult.getState() == SaveState.SAVING)) {
            // LOG.info("PS " + psAttemptId + " need save " + subSaveContext);
            resBuilder.setNeedSaveMatrices(ProtobufUtil.convert(subSaveContext));
        }
        // Check is there load request
        PSMatricesLoadContext subLoadContext = context.getModelLoader().getLoadContext(psAttemptId.getPsId());
        PSMatricesLoadResult subLoadResult = context.getModelLoader().getLoadResult(psAttemptId.getPsId());
        if (subLoadContext != null && subLoadResult != null && subLoadContext.getRequestId() == subLoadResult.getRequestId() && (subLoadResult.getState() == LoadState.INIT || subLoadResult.getState() == LoadState.LOADING)) {
            // LOG.info("PS " + psAttemptId + " need load " + subLoadContext);
            resBuilder.setNeedLoadMatrices(ProtobufUtil.convert(subLoadContext));
        }
        // check matrix metadata inconsistencies between master and parameter server.
        // if a matrix exists on the Master and does not exist on ps, then it is necessary to notify ps to establish the matrix
        // if a matrix exists on the ps and does not exist on master, then it is necessary to notify ps to remove the matrix
        List<MatrixReportProto> matrixReportsProto = request.getMatrixReportsList();
        List<Integer> needReleaseMatrices = new ArrayList<>();
        List<MatrixMeta> needCreateMatrices = new ArrayList<>();
        List<RecoverPartKey> needRecoverParts = new ArrayList<>();
        List<MatrixReport> matrixReports = ProtobufUtil.convertToMatrixReports(matrixReportsProto);
        context.getMatrixMetaManager().syncMatrixInfos(matrixReports, needCreateMatrices, needReleaseMatrices, needRecoverParts, psAttemptId.getPsId());
        size = needCreateMatrices.size();
        for (int i = 0; i < size; i++) {
            resBuilder.addNeedCreateMatrices(ProtobufUtil.convertToMatrixMetaProto(needCreateMatrices.get(i)));
        }
        size = needReleaseMatrices.size();
        for (int i = 0; i < size; i++) {
            resBuilder.addNeedReleaseMatrixIds(needReleaseMatrices.get(i));
        }
        size = needRecoverParts.size();
        for (int i = 0; i < size; i++) {
            resBuilder.addNeedRecoverParts(ProtobufUtil.convert(needRecoverParts.get(i)));
        }
    }
    return resBuilder.build();
}
Also used : MatrixReportProto(com.tencent.angel.protobuf.generated.PSMasterServiceProtos.MatrixReportProto) Int2IntOpenHashMap(it.unimi.dsi.fastutil.ints.Int2IntOpenHashMap) LinkedHashMap(java.util.LinkedHashMap) HashMap(java.util.HashMap) MatrixMeta(com.tencent.angel.ml.matrix.MatrixMeta) ArrayList(java.util.ArrayList) PSAttemptId(com.tencent.angel.ps.PSAttemptId) Pair(com.tencent.angel.protobuf.generated.MLProtos.Pair) RecoverPartKey(com.tencent.angel.ps.ha.RecoverPartKey) PSReportResponse(com.tencent.angel.protobuf.generated.PSMasterServiceProtos.PSReportResponse) PSAttemptStateUpdateEvent(com.tencent.angel.master.ps.attempt.PSAttemptStateUpdateEvent) PSMatricesLoadContext(com.tencent.angel.model.PSMatricesLoadContext) PSMatricesSaveContext(com.tencent.angel.model.PSMatricesSaveContext) PSMatricesSaveResult(com.tencent.angel.model.PSMatricesSaveResult) MatrixReport(com.tencent.angel.ml.matrix.MatrixReport) PSMatricesLoadResult(com.tencent.angel.model.PSMatricesLoadResult)

Example 2 with RecoverPartKey

use of com.tencent.angel.ps.ha.RecoverPartKey in project angel by Tencent.

the class AMMatrixMetaManager method syncMatrixInfos.

/**
 * compare the matrix meta on the master and the matrix meta on ps to find the matrix this
 * parameter server needs to create and delete
 *
 * @param matrixReports parameter server matrix report, include the matrix ids this parameter
 * server hold.
 * @param needCreateMatrixes use to return the matrix partitions this parameter server need to
 * build
 * @param needReleaseMatrixes use to return the matrix ids this parameter server need to remove
 * @param needRecoverParts need recover partitions
 * @param psId parameter server id
 */
public void syncMatrixInfos(List<MatrixReport> matrixReports, List<MatrixMeta> needCreateMatrixes, List<Integer> needReleaseMatrixes, List<RecoverPartKey> needRecoverParts, ParameterServerId psId) {
    // get matrix ids in the parameter server report
    IntOpenHashSet matrixInPS = new IntOpenHashSet();
    int size = matrixReports.size();
    for (int i = 0; i < size; i++) {
        matrixInPS.add(matrixReports.get(i).matrixId);
    }
    handleMatrixReports(psId, matrixReports);
    Set<RecoverPartKey> parts = getAndRemoveNeedRecoverParts(psId);
    if (parts != null) {
        needRecoverParts.addAll(parts);
    }
    // get the matrices parameter server need to create and delete
    getPSNeedUpdateMatrix(matrixInPS, needCreateMatrixes, needReleaseMatrixes, psId);
    psMatricesUpdate(psId, matrixReports);
}
Also used : IntOpenHashSet(it.unimi.dsi.fastutil.ints.IntOpenHashSet) RecoverPartKey(com.tencent.angel.ps.ha.RecoverPartKey)

Example 3 with RecoverPartKey

use of com.tencent.angel.ps.ha.RecoverPartKey in project angel by Tencent.

the class AMMatrixMetaManager method handlePartReport.

private void handlePartReport(ParameterServerId psId, int matrixId, PartReport partReport) {
    ParameterServerId master = matrixMetaManager.getMasterPs(matrixId, partReport.partId);
    if (!psId.equals(master)) {
        MatrixMeta matrixMeta = matrixMetaManager.getMatrixMeta(matrixId);
        if (matrixMeta == null) {
            return;
        }
        matrixMeta.getPartitionMeta(partReport.partId).addReplicationPS(psId);
        if (partReport.state == PartitionState.INITIALIZING) {
            addNeedRecoverPart(master, new RecoverPartKey(new PartitionKey(matrixId, partReport.partId), new PSLocation(psId, context.getLocationManager().getPsLocation(psId))));
        } else if (partReport.state == PartitionState.READ_AND_WRITE) {
            ParameterServerId orignalMaster = matrixPartitionsOnPS.get(psId).get(matrixId).getPartitionMeta(partReport.partId).getMasterPs();
            if (orignalMaster.equals(psId)) {
                matrixMetaManager.getMatrixMeta(matrixId).getPartitionMeta(partReport.partId).makePsToMaster(psId);
            }
        }
    }
}
Also used : RecoverPartKey(com.tencent.angel.ps.ha.RecoverPartKey) PSLocation(com.tencent.angel.ps.server.data.PSLocation) PartitionKey(com.tencent.angel.PartitionKey) ParameterServerId(com.tencent.angel.ps.ParameterServerId)

Aggregations

RecoverPartKey (com.tencent.angel.ps.ha.RecoverPartKey)3 PartitionKey (com.tencent.angel.PartitionKey)1 PSAttemptStateUpdateEvent (com.tencent.angel.master.ps.attempt.PSAttemptStateUpdateEvent)1 MatrixMeta (com.tencent.angel.ml.matrix.MatrixMeta)1 MatrixReport (com.tencent.angel.ml.matrix.MatrixReport)1 PSMatricesLoadContext (com.tencent.angel.model.PSMatricesLoadContext)1 PSMatricesLoadResult (com.tencent.angel.model.PSMatricesLoadResult)1 PSMatricesSaveContext (com.tencent.angel.model.PSMatricesSaveContext)1 PSMatricesSaveResult (com.tencent.angel.model.PSMatricesSaveResult)1 Pair (com.tencent.angel.protobuf.generated.MLProtos.Pair)1 MatrixReportProto (com.tencent.angel.protobuf.generated.PSMasterServiceProtos.MatrixReportProto)1 PSReportResponse (com.tencent.angel.protobuf.generated.PSMasterServiceProtos.PSReportResponse)1 PSAttemptId (com.tencent.angel.ps.PSAttemptId)1 ParameterServerId (com.tencent.angel.ps.ParameterServerId)1 PSLocation (com.tencent.angel.ps.server.data.PSLocation)1 Int2IntOpenHashMap (it.unimi.dsi.fastutil.ints.Int2IntOpenHashMap)1 IntOpenHashSet (it.unimi.dsi.fastutil.ints.IntOpenHashSet)1 ArrayList (java.util.ArrayList)1 HashMap (java.util.HashMap)1 LinkedHashMap (java.util.LinkedHashMap)1