@Override
public void process(Pair<TransactionTrace, Integer> p) {
assert (p != null);
final TransactionTrace txn_trace = p.getFirst();
final int i = p.getSecond(); // Interval
final int txn_weight = (use_txn_weights ? txn_trace.getWeight() : 1);
final String proc_key =
CatalogKey.createKey(CatalogUtil.DEFAULT_DATABASE_NAME, txn_trace.getCatalogItemName());
// Terrible Hack: Assume that we are using the SingleSitedCostModel
// and that
// it will return fixed values based on whether the txn is
// single-partitioned or not
SingleSitedCostModel singlesited_cost_model = (SingleSitedCostModel) cost_models[i];
total_interval_txns[i] += txn_weight;
total_interval_queries[i] += (txn_trace.getQueryCount() * txn_weight);
histogram_procs.put(proc_key, txn_weight);
try {
singlesited_cost_model.estimateTransactionCost(catalogContext, workload, filter, txn_trace);
TransactionCacheEntry txn_entry =
singlesited_cost_model.getTransactionCacheEntry(txn_trace);
assert (txn_entry != null) : "No txn entry for " + txn_trace;
Collection<Integer> partitions = txn_entry.getTouchedPartitions();
// If the txn runs on only one partition, then the cost is
// nothing
if (txn_entry.isSinglePartitioned()) {
singlepartition_ctrs[i] += txn_weight;
if (!partitions.isEmpty()) {
assert (txn_entry.getAllTouchedPartitionsHistogram().getValueCount() == 1)
: txn_entry
+ " says it was single-partitioned but the partition count says otherwise:\n"
+ txn_entry.debug();
singlepartition_with_partitions_ctrs[i] += txn_weight;
}
histogram_sp_procs.put(proc_key, txn_weight);
// If the txn runs on multiple partitions, then the cost
// is...
// XXX 2010-06-28: The number of partitions that the txn
// touches divided by the total number of partitions
// XXX 2010-07-02: The histogram for the total number of
// partitions touched by all of the queries
// in the transaction. This ensures that txns with just one
// multi-partition query
// isn't weighted the same as a txn with many
// multi-partition queries
} else {
assert (!partitions.isEmpty()) : "No touched partitions for " + txn_trace;
if (partitions.size() == 1
&& txn_entry.getExecutionPartition() != HStoreConstants.NULL_PARTITION_ID) {
assert (CollectionUtil.first(partitions) != txn_entry.getExecutionPartition())
: txn_entry.debug();
exec_mismatch_ctrs[i] += txn_weight;
partitions_touched[i] += txn_weight;
} else {
assert (partitions.size() > 1)
: String.format(
"%s is not marked as single-partition but it only touches one partition\n%s",
txn_trace, txn_entry.debug());
}
partitions_touched[i] += (partitions.size() * txn_weight); // Txns
multipartition_ctrs[i] += txn_weight;
histogram_mp_procs.put(proc_key, txn_weight);
}
Integer base_partition = txn_entry.getExecutionPartition();
if (base_partition != null) {
exec_histogram[i].put(base_partition, txn_weight);
} else {
exec_histogram[i].put(all_partitions, txn_weight);
}
if (debug.val) { // &&
// txn_trace.getCatalogItemName().equalsIgnoreCase("DeleteCallForwarding"))
// {
Procedure catalog_proc = txn_trace.getCatalogItem(catalogContext.database);
Map<String, Object> inner = new LinkedHashMap<String, Object>();
for (Statement catalog_stmt : catalog_proc.getStatements()) {
inner.put(catalog_stmt.fullName(), CatalogUtil.getReferencedTables(catalog_stmt));
}
Map<String, Object> m = new LinkedHashMap<String, Object>();
m.put(txn_trace.toString(), null);
m.put("Interval", i);
m.put("Single-Partition", txn_entry.isSinglePartitioned());
m.put("Base Partition", base_partition);
m.put("Touched Partitions", partitions);
m.put(catalog_proc.fullName(), inner);
LOG.debug(StringUtil.formatMaps(m));
}
// We need to keep a count of the number txns that didn't have
// all of its queries estimated
// completely so that we can update the access histograms down
// below for entropy calculations
// Note that this is at the txn level, not the query level.
if (!txn_entry.isComplete()) {
incomplete_txn_ctrs[i] += txn_weight;
tmp_missingPartitions.clear();
tmp_missingPartitions.addAll(all_partitions);
tmp_missingPartitions.removeAll(txn_entry.getTouchedPartitions());
// Update the histogram for this interval to keep track of
// how many times we need to
// increase the partition access histogram
incomplete_txn_histogram[i].put(tmp_missingPartitions, txn_weight);
if (trace.val) {
Map<String, Object> m = new LinkedHashMap<String, Object>();
m.put(String.format("Marking %s as incomplete in interval #%d", txn_trace, i), null);
m.put("Examined Queries", txn_entry.getExaminedQueryCount());
m.put("Total Queries", txn_entry.getTotalQueryCount());
m.put("Touched Partitions", txn_entry.getTouchedPartitions());
m.put("Missing Partitions", tmp_missingPartitions);
LOG.trace(StringUtil.formatMaps(m));
}
}
} catch (Exception ex) {
CatalogUtil.saveCatalog(catalogContext.catalog, CatalogUtil.CATALOG_FILENAME);
throw new RuntimeException(
"Failed to estimate cost for " + txn_trace.getCatalogItemName() + " at interval " + i,
ex);
}
}
@Override
protected double estimateWorkloadCostImpl(
final CatalogContext catalogContext,
final Workload workload,
final Filter filter,
final Double upper_bound)
throws Exception {
if (debug.val)
LOG.debug(
"Calculating workload execution cost across "
+ num_intervals
+ " intervals for "
+ num_partitions
+ " partitions");
// (1) Grab the costs at the different time intervals
// Also create the ratios that we will use to weight the interval costs
final AtomicLong total_txns = new AtomicLong(0);
// final HashSet<Long> trace_ids[] = new HashSet[num_intervals];
for (int i = 0; i < num_intervals; i++) {
total_interval_txns[i] = 0;
total_interval_queries[i] = 0;
singlepartition_ctrs[i] = 0;
singlepartition_with_partitions_ctrs[i] = 0;
multipartition_ctrs[i] = 0;
partitions_touched[i] = 0;
incomplete_txn_ctrs[i] = 0;
exec_mismatch_ctrs[i] = 0;
incomplete_txn_histogram[i].clear();
missing_txn_histogram[i].clear();
exec_histogram[i].clear();
} // FOR
// (2) Now go through the workload and estimate the partitions that each txn
// will touch for the given catalog setups
if (trace.val) {
LOG.trace("Total # of Txns in Workload: " + workload.getTransactionCount());
if (filter != null)
LOG.trace(
"Workload Filter Chain: " + StringUtil.join(" ", "\n", filter.getFilters()));
}
// QUEUING THREAD
tmp_consumers.clear();
Producer<TransactionTrace, Pair<TransactionTrace, Integer>> producer =
new Producer<TransactionTrace, Pair<TransactionTrace, Integer>>(
CollectionUtil.iterable(workload.iterator(filter))) {
@Override
public Pair<Consumer<Pair<TransactionTrace, Integer>>, Pair<TransactionTrace, Integer>>
transform(TransactionTrace txn_trace) {
int i = workload.getTimeInterval(txn_trace, num_intervals);
assert (i >= 0)
: "Invalid time interval '" + i + "'\n" + txn_trace.debug(catalogContext.database);
assert (i < num_intervals)
: "Invalid interval: " + i + "\n" + txn_trace.debug(catalogContext.database);
total_txns.incrementAndGet();
Pair<TransactionTrace, Integer> p = Pair.of(txn_trace, i);
return (Pair.of(tmp_consumers.get(i), p));
}
};
// PROCESSING THREADS
final int num_threads = ThreadUtil.getMaxGlobalThreads();
int interval_ctr = 0;
for (int thread = 0; thread < num_threads; thread++) {
// First create a new IntervalProcessor/Consumer
IntervalProcessor ip = new IntervalProcessor(catalogContext, workload, filter);
// Then assign it to some number of intervals
for (int i = 0, cnt = (int) Math.ceil(num_intervals / (double) num_threads); i < cnt; i++) {
if (interval_ctr > num_intervals) break;
tmp_consumers.put(interval_ctr++, ip);
if (trace.val)
LOG.trace(
String.format("Interval #%02d => IntervalProcessor #%02d", interval_ctr - 1, thread));
} // FOR
// And make sure that we queue it up too
producer.addConsumer(ip);
} // FOR (threads)
ThreadUtil.runGlobalPool(producer.getRunnablesList()); // BLOCKING
if (debug.val) {
int processed = 0;
for (Consumer<?> c : producer.getConsumers()) {
processed += c.getProcessedCounter();
} // FOR
assert (total_txns.get() == processed)
: String.format("Expected[%d] != Processed[%d]", total_txns.get(), processed);
}
// We have to convert all of the costs into the range of [0.0, 1.0]
// For each interval, divide the number of partitions touched by the total number
// of partitions that the interval could have touched (worst case scenario)
final double execution_costs[] = new double[num_intervals];
StringBuilder sb = (this.isDebugEnabled() || debug.get() ? new StringBuilder() : null);
Map<String, Object> debug_m = null;
if (sb != null) {
debug_m = new LinkedHashMap<String, Object>();
}
if (debug.val)
LOG.debug("Calculating execution cost for " + this.num_intervals + " intervals...");
long total_multipartition_txns = 0;
for (int i = 0; i < this.num_intervals; i++) {
interval_weights[i] = total_interval_txns[i] / (double) total_txns.get();
long total_txns_in_interval = (long) total_interval_txns[i];
long total_queries_in_interval = (long) total_interval_queries[i];
long num_txns = this.cost_models[i].txn_ctr.get();
long potential_txn_touches = (total_txns_in_interval * num_partitions); // TXNS
double penalty = 0.0d;
total_multipartition_txns += multipartition_ctrs[i];
// Divide the total number of partitions touched by...
// This is the total number of partitions that we could have touched
// in this interval
// And this is the total number of partitions that we did actually touch
if (multipartition_ctrs[i] > 0) {
assert (partitions_touched[i] > 0) : "No touched partitions for interval " + i;
double cost = (partitions_touched[i] / (double) potential_txn_touches);
if (this.use_multitpartition_penalty) {
penalty =
this.multipartition_penalty
* (1.0d + (multipartition_ctrs[i] / (double) total_txns_in_interval));
assert (penalty >= 1.0) : "The multipartition penalty is less than one: " + penalty;
cost *= penalty;
}
execution_costs[i] = Math.min(cost, (double) potential_txn_touches);
}
// For each txn that wasn't even evaluated, add all of the
// partitions to the incomplete histogram
if (num_txns < total_txns_in_interval) {
if (trace.val)
LOG.trace(
"Adding "
+ (total_txns_in_interval - num_txns)
+ " entries to the incomplete histogram for interval #"
+ i);
for (long ii = num_txns; ii < total_txns_in_interval; ii++) {
missing_txn_histogram[i].put(all_partitions);
} // WHILE
}
if (sb != null) {
tmp_penalties.add(penalty);
tmp_total.add(total_txns_in_interval);
tmp_touched.add(partitions_touched[i]);
tmp_potential.add(potential_txn_touches);
Map<String, Object> inner = new LinkedHashMap<String, Object>();
inner.put("Partitions Touched", partitions_touched[i]);
inner.put("Potential Touched", potential_txn_touches);
inner.put("Multi-Partition Txns", multipartition_ctrs[i]);
inner.put("Total Txns", total_txns_in_interval);
inner.put("Total Queries", total_queries_in_interval);
inner.put("Missing Txns", (total_txns_in_interval - num_txns));
inner.put("Cost", String.format("%.05f", execution_costs[i]));
inner.put("Exec Txns", exec_histogram[i].getSampleCount());
debug_m.put("Interval #" + i, inner);
}
} // FOR
if (sb != null) {
Map<String, Object> m0 = new LinkedHashMap<String, Object>();
m0.put("SinglePartition Txns", (total_txns.get() - total_multipartition_txns));
m0.put("MultiPartition Txns", total_multipartition_txns);
m0.put(
"Total Txns",
String.format(
"%d [%.06f]",
total_txns.get(), (1.0d - (total_multipartition_txns / (double) total_txns.get()))));
Map<String, Object> m1 = new LinkedHashMap<String, Object>();
m1.put("Touched Partitions", tmp_touched);
m1.put("Potential Partitions", tmp_potential);
m1.put("Total Partitions", tmp_total);
m1.put("Penalties", tmp_penalties);
sb.append(StringUtil.formatMaps(debug_m, m0, m1));
if (debug.val) LOG.debug("**** Execution Cost ****\n" + sb);
this.appendDebugMessage(sb);
}
// LOG.debug("Execution By Intervals:\n" + sb.toString());
// (3) We then need to go through and grab the histograms of partitions were accessed
if (sb != null) {
if (debug.val)
LOG.debug("Calculating skew factor for " + this.num_intervals + " intervals...");
debug_histograms.clear();
sb = new StringBuilder();
}
for (int i = 0; i < this.num_intervals; i++) {
ObjectHistogram<Integer> histogram_txn = this.cost_models[i].getTxnPartitionAccessHistogram();
ObjectHistogram<Integer> histogram_query =
this.cost_models[i].getQueryPartitionAccessHistogram();
this.histogram_query_partitions.put(histogram_query);
long num_queries = this.cost_models[i].query_ctr.get();
this.query_ctr.addAndGet(num_queries);
// DEBUG
SingleSitedCostModel inner_costModel = (SingleSitedCostModel) this.cost_models[i];
boolean is_valid =
(partitions_touched[i] + singlepartition_with_partitions_ctrs[i])
== (this.cost_models[i].getTxnPartitionAccessHistogram().getSampleCount()
+ exec_mismatch_ctrs[i]);
if (!is_valid) {
LOG.error("Transaction Entries: " + inner_costModel.getTransactionCacheEntries().size());
ObjectHistogram<Integer> check = new ObjectHistogram<Integer>();
for (TransactionCacheEntry tce : inner_costModel.getTransactionCacheEntries()) {
check.put(tce.getTouchedPartitions());
// LOG.error(tce.debug() + "\n");
}
LOG.error(
"Check Touched Partitions: sample="
+ check.getSampleCount()
+ ", values="
+ check.getValueCount());
LOG.error(
"Cache Touched Partitions: sample="
+ this.cost_models[i].getTxnPartitionAccessHistogram().getSampleCount()
+ ", values="
+ this.cost_models[i].getTxnPartitionAccessHistogram().getValueCount());
int qtotal = inner_costModel.getAllQueryCacheEntries().size();
int ctr = 0;
int multip = 0;
for (QueryCacheEntry qce : inner_costModel.getAllQueryCacheEntries()) {
ctr += (qce.getAllPartitions().isEmpty() ? 0 : 1);
multip += (qce.getAllPartitions().size() > 1 ? 1 : 0);
} // FOR
LOG.error("# of QueryCacheEntries with Touched Partitions: " + ctr + " / " + qtotal);
LOG.error("# of MultiP QueryCacheEntries: " + multip);
}
assert (is_valid)
: String.format(
"Partitions Touched by Txns Mismatch in Interval #%d\n"
+ "(partitions_touched[%d] + singlepartition_with_partitions_ctrs[%d]) != "
+ "(histogram_txn[%d] + exec_mismatch_ctrs[%d])",
i,
partitions_touched[i],
singlepartition_with_partitions_ctrs[i],
this.cost_models[i].getTxnPartitionAccessHistogram().getSampleCount(),
exec_mismatch_ctrs[i]);
this.histogram_java_partitions.put(this.cost_models[i].getJavaExecutionHistogram());
this.histogram_txn_partitions.put(histogram_txn);
long num_txns = this.cost_models[i].txn_ctr.get();
assert (num_txns >= 0) : "The transaction counter at interval #" + i + " is " + num_txns;
this.txn_ctr.addAndGet(num_txns);
// Calculate the skew factor at this time interval
// XXX: Should the number of txns be the total number of unique txns
// that were executed or the total number of times a txn touched the partitions?
// XXX: What do we do when the number of elements that we are examining is zero?
// I guess the cost just needs to be zero?
// XXX: What histogram do we want to use?
target_histogram.clear();
target_histogram.put(histogram_txn);
// For each txn that we haven't gotten an estimate for at this interval,
// we're going mark it as being broadcast to all partitions. That way the access
// histogram will look uniform. Then as more information is added, we will
// This is an attempt to make sure that the skew cost never decreases but only increases
long total_txns_in_interval = (long) total_interval_txns[i];
if (sb != null) {
debug_histograms.put("Incomplete Txns", incomplete_txn_histogram[i]);
debug_histograms.put("Missing Txns", missing_txn_histogram[i]);
debug_histograms.put(
"Target Partitions (BEFORE)", new ObjectHistogram<Integer>(target_histogram));
debug_histograms.put("Target Partitions (AFTER)", target_histogram);
}
// Merge the values from incomplete histogram into the target
// histogram
target_histogram.put(incomplete_txn_histogram[i]);
target_histogram.put(missing_txn_histogram[i]);
exec_histogram[i].put(missing_txn_histogram[i]);
long num_elements = target_histogram.getSampleCount();
// The number of partition touches should never be greater than our
// potential touches
assert (num_elements <= (total_txns_in_interval * num_partitions))
: "New Partitions Touched Sample Count ["
+ num_elements
+ "] < "
+ "Maximum Potential Touched Count ["
+ (total_txns_in_interval * num_partitions)
+ "]";
if (sb != null) {
Map<String, Object> m = new LinkedHashMap<String, Object>();
for (String key : debug_histograms.keySet()) {
ObjectHistogram<?> h = debug_histograms.get(key);
m.put(
key,
String.format("[Sample=%d, Value=%d]\n%s", h.getSampleCount(), h.getValueCount(), h));
} // FOR
sb.append(
String.format(
"INTERVAL #%d [total_txns_in_interval=%d, num_txns=%d, incomplete_txns=%d]\n%s",
i,
total_txns_in_interval,
num_txns,
incomplete_txn_ctrs[i],
StringUtil.formatMaps(m)));
}
// Txn Skew
if (num_elements == 0) {
txn_skews[i] = 0.0d;
} else {
txn_skews[i] = SkewFactorUtil.calculateSkew(num_partitions, num_elements, target_histogram);
}
// Exec Skew
if (exec_histogram[i].getSampleCount() == 0) {
exec_skews[i] = 0.0d;
} else {
exec_skews[i] =
SkewFactorUtil.calculateSkew(
num_partitions, exec_histogram[i].getSampleCount(), exec_histogram[i]);
}
total_skews[i] = (0.5 * exec_skews[i]) + (0.5 * txn_skews[i]);
if (sb != null) {
sb.append("Txn Skew = " + MathUtil.roundToDecimals(txn_skews[i], 6) + "\n");
sb.append("Exec Skew = " + MathUtil.roundToDecimals(exec_skews[i], 6) + "\n");
sb.append("Total Skew = " + MathUtil.roundToDecimals(total_skews[i], 6) + "\n");
sb.append(StringUtil.DOUBLE_LINE);
}
} // FOR
if (sb != null && sb.length() > 0) {
if (debug.val) LOG.debug("**** Skew Factor ****\n" + sb);
this.appendDebugMessage(sb);
}
if (trace.val) {
for (int i = 0; i < num_intervals; i++) {
LOG.trace(
"Time Interval #"
+ i
+ "\n"
+ "Total # of Txns: "
+ this.cost_models[i].txn_ctr.get()
+ "\n"
+ "Multi-Partition Txns: "
+ multipartition_ctrs[i]
+ "\n"
+ "Execution Cost: "
+ execution_costs[i]
+ "\n"
+ "ProcHistogram:\n"
+ this.cost_models[i].getProcedureHistogram().toString()
+ "\n"
+
// "TransactionsPerPartitionHistogram:\n" +
// this.cost_models[i].getTxnPartitionAccessHistogram()
// + "\n" +
StringUtil.SINGLE_LINE);
}
}
// (3) We can now calculate the final total estimate cost of this workload as the following
// Just take the simple ratio of mp txns / all txns
this.last_execution_cost =
MathUtil.weightedMean(
execution_costs,
total_interval_txns); // MathUtil.roundToDecimals(MathUtil.geometricMean(execution_costs,
// MathUtil.GEOMETRIC_MEAN_ZERO),
// 10);
// The final skew cost needs to be weighted by the percentage of txns running in that interval
// This will cause the partitions with few txns
this.last_skew_cost =
MathUtil.weightedMean(
total_skews, total_interval_txns); // roundToDecimals(MathUtil.geometricMean(entropies,
// MathUtil.GEOMETRIC_MEAN_ZERO),
// 10);
double new_final_cost =
(this.use_execution ? (this.execution_weight * this.last_execution_cost) : 0)
+ (this.use_skew ? (this.skew_weight * this.last_skew_cost) : 0);
if (sb != null) {
Map<String, Object> m = new LinkedHashMap<String, Object>();
m.put("Total Txns", total_txns.get());
m.put("Interval Txns", Arrays.toString(total_interval_txns));
m.put("Execution Costs", Arrays.toString(execution_costs));
m.put("Skew Factors", Arrays.toString(total_skews));
m.put("Txn Skew", Arrays.toString(txn_skews));
m.put("Exec Skew", Arrays.toString(exec_skews));
m.put("Interval Weights", Arrays.toString(interval_weights));
m.put(
"Final Cost",
String.format(
"%f = %f + %f", new_final_cost, this.last_execution_cost, this.last_skew_cost));
if (debug.val) LOG.debug(StringUtil.formatMaps(m));
this.appendDebugMessage(StringUtil.formatMaps(m));
}
this.last_final_cost = new_final_cost;
return (MathUtil.roundToDecimals(this.last_final_cost, 5));
}