void
unpackMetrics(Prof::CallPath::Profile& profile,
const ParallelAnalysis::PackedMetrics& packedMetrics)
{
Prof::CCT::Tree& cct = *profile.cct();
// 1. unpack 'packedMetrics' into temporary derived metrics [mBegId,
// mEndId) in 'profile'
uint mBegId = packedMetrics.mBegId(), mEndId = packedMetrics.mEndId();
DIAG_Assert(packedMetrics.numNodes() == cct.maxDenseId() + 1, "");
DIAG_Assert(packedMetrics.numMetrics() == mEndId - mBegId, "");
for (uint nodeId = 1; nodeId < packedMetrics.numNodes(); ++nodeId) {
for (uint mId1 = 0, mId2 = mBegId; mId2 < mEndId; ++mId1, ++mId2) {
Prof::CCT::ANode* n = cct.findNode(nodeId);
n->demandMetric(mId2) = packedMetrics.idx(nodeId, mId1);
}
}
// 2. update derived metrics [mDrvdBeg, mDrvdEnd) based on new
// values in [mBegId, mEndId)
uint mDrvdBeg = packedMetrics.mDrvdBegId();
uint mDrvdEnd = packedMetrics.mDrvdEndId();
cct.root()->computeMetricsIncr(*profile.metricMgr(), mDrvdBeg, mDrvdEnd,
Prof::Metric::AExprIncr::FnCombine);
}
// mergeCilkMain: cilk_main is called from two distinct call sites
// within the runtime, resulting in an undesirable bifurcation within
// the CCT. The easiest way to fix this is to use a normalization
// step.
static void
mergeCilkMain(Prof::CallPath::Profile& prof)
{
using namespace Prof;
CCT::ProcFrm* mainFrm = NULL;
// 1. attempt to find 'CilkNameMgr::cilkmain'
for (CCT::ANodeIterator it(prof.cct()->root(),
&CCT::ANodeTyFilter[CCT::ANode::TyProcFrm]);
it.Current(); ++it) {
CCT::ProcFrm* x = static_cast<CCT::ProcFrm*>(it.current());
if (x->procName() == CilkNameMgr::cilkmain) {
mainFrm = x;
break;
}
}
// 2. merge any sibling 'CilkNameMgr::cilkmain'
if (mainFrm) {
CCT::ANodeChildIterator it(mainFrm->parent(),
&CCT::ANodeTyFilter[CCT::ANode::TyProcFrm]);
for ( ; it.Current(); /* */) {
CCT::ProcFrm* x = static_cast<CCT::ProcFrm*>(it.current());
it++; // advance iterator -- it is pointing at 'x'
if (x->procName() == CilkNameMgr::cilkmain) {
mainFrm->merge(x); // deletes 'x'
}
}
}
}
void
Analysis::CallPath::
noteStaticStructureOnLeaves(Prof::CallPath::Profile& prof)
{
const Prof::Struct::Root* rootStrct = prof.structure()->root();
Prof::CCT::ANodeIterator it(prof.cct()->root(), NULL/*filter*/,
true/*leavesOnly*/, IteratorStack::PreOrder);
for (Prof::CCT::ANode* n = NULL; (n = it.current()); ++it) {
Prof::CCT::ADynNode* n_dyn = dynamic_cast<Prof::CCT::ADynNode*>(n);
if (n_dyn) {
Prof::LoadMap::LMId_t lmId = n_dyn->lmId(); // ok if LoadMap::LMId_NULL
Prof::LoadMap::LM* loadmap_lm = prof.loadmap()->lm(lmId);
const string& lm_nm = loadmap_lm->name();
const Prof::Struct::LM* lmStrct = rootStrct->findLM(lm_nm);
DIAG_Assert(lmStrct, "failed to find Struct::LM: " << lm_nm);
VMA lm_ip = n_dyn->lmIP();
const Prof::Struct::ACodeNode* strct = lmStrct->findByVMA(lm_ip);
DIAG_Assert(strct, "Analysis::CallPath::noteStaticStructureOnLeaves: failed to find structure for: " << n_dyn->toStringMe(Prof::CCT::Tree::OFlg_DebugAll));
n->structure(strct);
}
}
}
// overlayStaticStructure: Create frames for CCT::Call and CCT::Stmt
// nodes using a preorder walk over the CCT.
void
Analysis::CallPath::
overlayStaticStructure(Prof::CallPath::Profile& prof,
Prof::LoadMap::LM* loadmap_lm,
Prof::Struct::LM* lmStrct, BinUtil::LM* lm)
{
overlayStaticStructure(prof.cct()->root(), loadmap_lm, lmStrct, lm);
}
void
Analysis::CallPath::pruneBySummaryMetrics(Prof::CallPath::Profile& prof,
uint8_t* prunedNodes)
{
VMAIntervalSet ivalset;
const Prof::Metric::Mgr& mMgrGbl = *(prof.metricMgr());
for (uint mId = 0; mId < mMgrGbl.size(); ++mId) {
const Prof::Metric::ADesc* m = mMgrGbl.metric(mId);
if (m->isVisible()
&& m->type() == Prof::Metric::ADesc::TyIncl
&& (m->nameBase().find("Sum") != string::npos)) {
ivalset.insert(VMAInterval(mId, mId + 1)); // [ )
}
}
prof.cct()->root()->pruneByMetrics(*prof.metricMgr(), ivalset,
prof.cct()->root(), 0.001,
prunedNodes);
}
static void
noteStaticStructure(Prof::CallPath::Profile& prof)
{
using namespace Prof;
const Prof::CCT::ANode* cct_root = prof.cct()->root();
for (CCT::ANodeIterator it(cct_root); it.Current(); ++it) {
CCT::ANode* x = it.current();
Prof::Struct::ACodeNode* strct = x->structure();
if (strct) {
strct->demandMetric(CallPath::Profile::StructMetricIdFlg) += 1.0;
}
}
}
// Assumes: metrics are of type Metric::SampledDesc and values are
// only at leaves (CCT::Stmt)
void
MetricComponentsFact::make(Prof::CallPath::Profile& prof)
{
using namespace Prof;
// ------------------------------------------------------------
// Create destination metric descriptors and mapping from source
// metrics to destination metrics
// ------------------------------------------------------------
std::vector<uint> metricSrcIds;
std::vector<uint> metricDstIds;
Metric::Mgr* metricMgr = prof.metricMgr();
uint numMetrics_orig = metricMgr->size();
for (uint mId = 0; mId < numMetrics_orig; ++mId) {
Metric::ADesc* m = metricMgr->metric(mId);
if (MetricComponentsFact::isTimeMetric(m)) {
DIAG_Assert(typeid(*m) == typeid(Metric::SampledDesc), DIAG_UnexpectedInput << "temporary sanity check");
MetricComponentsFact::convertToWorkMetric(m);
metricSrcIds.push_back(m->id());
Metric::ADesc* m_new = m->clone();
m_new->nameBase("overhead");
m_new->description("parallel overhead");
metricMgr->insert(m_new);
DIAG_Assert(m_new->id() >= numMetrics_orig, "Currently, we assume new metrics are added at the end of the metric vector.");
metricDstIds.push_back(m_new->id());
}
}
if (metricSrcIds.empty()) {
return;
}
// ------------------------------------------------------------
// Create values for metric components
// ------------------------------------------------------------
make(prof.cct()->root(), metricSrcIds, metricDstIds, false);
}
void
packMetrics(const Prof::CallPath::Profile& profile,
ParallelAnalysis::PackedMetrics& packedMetrics)
{
Prof::CCT::Tree& cct = *profile.cct();
// pack derived metrics [mDrvdBeg, mDrvdEnd) from 'profile' into
// 'packedMetrics'
uint mDrvdBeg = packedMetrics.mDrvdBegId();
uint mDrvdEnd = packedMetrics.mDrvdEndId();
DIAG_Assert(packedMetrics.numNodes() == cct.maxDenseId() + 1, "");
DIAG_Assert(packedMetrics.numMetrics() == mDrvdEnd - mDrvdBeg, "");
for (Prof::CCT::ANodeIterator it(cct.root()); it.Current(); ++it) {
Prof::CCT::ANode* n = it.current();
for (uint mId1 = 0, mId2 = mDrvdBeg; mId2 < mDrvdEnd; ++mId1, ++mId2) {
packedMetrics.idx(n->id(), mId1) = n->metric(mId2);
}
}
}
// makeReturnCountMetric: A return count refers to the number of times
// a given CCT node is called by its parent context. However, when
// hpcrun records return counts, there is no structure (e.g. procedure
// frames) in the CCT. An an example, in the CCT fragment below, the
// return count [3] at 0xc means that 0xc returned to 0xbeef 3 times.
// Simlarly, 0xbeef returned to its caller 5 times.
//
// | |
// ip: 0xbeef [5] |
// / | \ |
// 0xa [1] 0xb [2] 0xc [3] |
// | | | |
//
// To be able to say procedure F is called by procedure G x times
// within this context, it is necessary to aggregate these counts at
// the newly added procedure frames (Struct::ProcFrm).
static void
makeReturnCountMetric(Prof::CallPath::Profile& prof)
{
std::vector<uint> retCntId;
// -------------------------------------------------------
// find return count metrics, if any
// -------------------------------------------------------
Prof::Metric::Mgr* metricMgr = prof.metricMgr();
for (uint i = 0; i < metricMgr->size(); ++i) {
Prof::Metric::ADesc* m = metricMgr->metric(i);
if (m->nameBase().find(HPCRUN_METRIC_RetCnt) != string::npos) {
retCntId.push_back(m->id());
m->computedType(Prof::Metric::ADesc::ComputedTy_Final);
m->type(Prof::Metric::ADesc::TyExcl);
}
}
if (retCntId.empty()) {
return;
}
// -------------------------------------------------------
// propagate and aggregate return counts
// -------------------------------------------------------
Prof::CCT::ANode* cct_root = prof.cct()->root();
Prof::CCT::ANodeIterator it(cct_root, NULL/*filter*/, false/*leavesOnly*/,
IteratorStack::PostOrder);
for (Prof::CCT::ANode* n = NULL; (n = it.current()); ++it) {
if (typeid(*n) != typeid(Prof::CCT::ProcFrm) && n != cct_root) {
Prof::CCT::ANode* n_parent = n->parent();
for (uint i = 0; i < retCntId.size(); ++i) {
uint mId = retCntId[i];
n_parent->demandMetric(mId) += n->demandMetric(mId);
n->metric(mId) = 0.0;
}
}
}
}
static void
pruneTrivialNodes(Prof::CallPath::Profile& prof)
{
pruneTrivialNodes(prof.cct()->root());
}
static void
coalesceStmts(Prof::CallPath::Profile& prof)
{
coalesceStmts(prof.cct()->root());
}
// make: ...temporary holding pattern...
void
MPIBlameShiftIdlenessFact::make(Prof::CallPath::Profile& prof)
{
using namespace Prof;
// ------------------------------------------------------------
// Create destination metric descriptors and mapping from source
// metrics to destination metrics
// ------------------------------------------------------------
std::vector<uint> metricSrcIds;
std::vector<uint> metricBalanceIds;
std::vector<uint> metricImbalInclIds, metricImbalExclIds;
std::vector<uint> metricIdleInclIds;
Metric::Mgr* metricMgr = prof.metricMgr();
uint numMetrics_orig = metricMgr->size();
for (uint mId = 0; mId < numMetrics_orig; ++mId) {
Metric::ADesc* m = metricMgr->metric(mId);
// find main source metric
if (MetricComponentsFact::isTimeMetric(m)
&& MetricComponentsFact::isDerivedMetric(m, s_sum)
&& m->type() == Metric::ADesc::TyIncl
&& m->isVisible() /* not a temporary */) {
DIAG_Assert(m->computedType() == Prof::Metric::ADesc::ComputedTy_NonFinal,
DIAG_UnexpectedInput);
metricSrcIds.push_back(m->id());
// FIXME: For now we use only Metric::ADesc::DerivedIncrDesc()
// We should also support Metric::ADesc::DerivedDesc()
DIAG_Assert(typeid(*m) == typeid(Metric::DerivedIncrDesc), DIAG_UnexpectedInput);
Metric::DerivedIncrDesc* m_imbalIncl =
static_cast<Metric::DerivedIncrDesc*>(m->clone());
m_imbalIncl->nameBase("imbalance" + s_sum);
m_imbalIncl->description("imbalance for MPI SPMD executions");
m_imbalIncl->expr(new Metric::SumIncr(Metric::IData::npos, // FIXME:Sum
Metric::IData::npos));
Metric::DerivedIncrDesc* m_imbalExcl =
static_cast<Metric::DerivedIncrDesc*>(m_imbalIncl->clone());
m_imbalExcl->type(Metric::ADesc::TyExcl);
m_imbalExcl->expr(new Metric::SumIncr(Metric::IData::npos,
Metric::IData::npos));
m_imbalIncl->partner(m_imbalExcl);
m_imbalExcl->partner(m_imbalIncl);
Metric::DerivedIncrDesc* m_idleIncl =
static_cast<Metric::DerivedIncrDesc*>(m->clone());
m_idleIncl->nameBase("idleness" + s_sum);
m_idleIncl->description("idleness for MPI executions");
m_idleIncl->partner(NULL);
m_idleIncl->expr(new Metric::SumIncr(Metric::IData::npos, // FIXME:Sum
Metric::IData::npos));
metricMgr->insert(m_imbalIncl);
metricMgr->insert(m_imbalExcl);
metricMgr->insert(m_idleIncl);
m_imbalIncl->expr()->accumId(m_imbalIncl->id());
m_imbalExcl->expr()->accumId(m_imbalExcl->id());
m_idleIncl->expr()->accumId(m_idleIncl->id());
DIAG_Assert(m_imbalIncl->id() >= numMetrics_orig && m_imbalExcl->id() >= numMetrics_orig, "Currently, we assume new metrics are added at the end of the metric vector.");
metricImbalInclIds.push_back(m_imbalIncl->id());
metricImbalExclIds.push_back(m_imbalExcl->id());
metricIdleInclIds.push_back(m_idleIncl->id());
}
// find secondary source metric
if (MetricComponentsFact::isTimeMetric(m)
&& MetricComponentsFact::isDerivedMetric(m, s_cfvar)
&& m->type() == Metric::ADesc::TyIncl
&& m->isVisible() /* not a temporary */) {
DIAG_Assert(m->computedType() == Prof::Metric::ADesc::ComputedTy_NonFinal,
DIAG_UnexpectedInput);
metricBalanceIds.push_back(m->id());
}
}
DIAG_Assert(metricSrcIds.size() == metricBalanceIds.size(), DIAG_UnexpectedInput);
if (metricSrcIds.empty()) {
return;
}
// ------------------------------------------------------------
// Create values for metric components
// ------------------------------------------------------------
// Note that metrics are non-finalized!
CCT::ANode* cctRoot = prof.cct()->root();
uint metricBalancedId = metricBalanceIds[0];
Metric::AExprIncr* metricBalancedExpr = dynamic_cast<Metric::DerivedIncrDesc*>(metricMgr->metric(metricBalancedId))->expr();
Metric::IData cctRoot_mdata(*cctRoot);
metricBalancedExpr->finalize(cctRoot_mdata);
double balancedThreshold = 1.2 * cctRoot_mdata.demandMetric(metricBalancedId);
makeMetrics(cctRoot, metricSrcIds,
metricImbalInclIds, metricImbalExclIds, metricIdleInclIds,
metricBalancedId, metricBalancedExpr, balancedThreshold,
NULL, NULL);
VMAIntervalSet metricDstInclIdSet;
for (uint i = 0; i < metricImbalInclIds.size(); ++i) {
uint mId = metricImbalInclIds[i];
metricDstInclIdSet.insert(VMAInterval(mId, mId + 1)); // [ )
mId = metricIdleInclIds[i];
metricDstInclIdSet.insert(VMAInterval(mId, mId + 1)); // [ )
}
cctRoot->aggregateMetricsIncl(metricDstInclIdSet);
}