better mempool queueing

consensus/reactor.go

@@ -22,8 +22,6 @@ const (
 	DataChannel  = byte(0x21)
 	VoteChannel  = byte(0x22)
 
-	PeerStateKey = "ConsensusReactor.peerState"
-
 	peerGossipSleepDuration = 100 * time.Millisecond // Time to sleep if there's nothing to send.
 )
 
@@ -107,7 +105,7 @@ func (conR *ConsensusReactor) AddPeer(peer *p2p.Peer) {
 
 	// Create peerState for peer
 	peerState := NewPeerState(peer)
-	peer.Data.Set(PeerStateKey, peerState)
+	peer.Data.Set(types.PeerStateKey, peerState)
 
 	// Begin gossip routines for this peer.
 	go conR.gossipDataRoutine(peer, peerState)
@@ -138,7 +136,7 @@ func (conR *ConsensusReactor) Receive(chID byte, peer *p2p.Peer, msgBytes []byte
 	}
 
 	// Get peer states
-	ps := peer.Data.Get(PeerStateKey).(*PeerState)
+	ps := peer.Data.Get(types.PeerStateKey).(*PeerState)
 	_, msg, err := DecodeMessage(msgBytes)
 	if err != nil {
 		log.Warn("Error decoding message", "channel", chID, "peer", peer, "msg", msg, "error", err, "bytes", msgBytes)
@@ -588,6 +586,14 @@ func (ps *PeerState) GetRoundState() *PeerRoundState {
 	return &prs
 }
 
+// Returns an atomic snapshot of the PeerRoundState's height
+// used by the mempool to ensure peers are caught up before broadcasting new txs
+func (ps *PeerState) GetHeight() int {
+	ps.mtx.Lock()
+	defer ps.mtx.Unlock()
+	return ps.PeerRoundState.Height
+}
+
 func (ps *PeerState) SetHasProposal(proposal *types.Proposal) {
 	ps.mtx.Lock()
 	defer ps.mtx.Unlock()

consensus/state.go

@@ -657,6 +657,10 @@ func (cs *ConsensusState) createProposalBlock() (block *types.Block, blockParts
 		return
 	}
 	txs := cs.mempoolReactor.Mempool.GetProposalTxs()
+	MaxTxsPerBlock := 100 // TODO
+	if len(txs) > MaxTxsPerBlock {
+		txs = txs[:MaxTxsPerBlock]
+	}
 	block = &types.Block{
 		Header: &types.Header{
 			ChainID:        cs.state.ChainID,

mempool/mempool.go

@@ -19,7 +19,9 @@ type Mempool struct {
 	mtx   sync.Mutex
 	state *sm.State
 	cache *sm.BlockCache
-	txs   []types.Tx
+	txs   []types.Tx // TODO: we need to add a map to facilitate replace-by-fee
+
+	resetInfo ResetInfo // so broadcast routines can respond to mempool flushing
 }
 
 func NewMempool(state *sm.State) *Mempool {
@@ -59,6 +61,18 @@ func (mem *Mempool) GetProposalTxs() []types.Tx {
 	return mem.txs
 }
 
+// We use this to inform peer routines of how the mempool has been updated
+type ResetInfo struct {
+	Height   int
+	Included []Range
+	Invalid  []Range
+}
+
+type Range struct {
+	Start  int
+	Length int
+}
+
 // "block" is the new block being committed.
 // "state" is the result of state.AppendBlock("block").
 // Txs that are present in "block" are discarded from mempool.
@@ -75,33 +89,51 @@ func (mem *Mempool) ResetForBlockAndState(block *types.Block, state *sm.State) {
 		blockTxsMap[string(types.TxID(state.ChainID, tx))] = struct{}{}
 	}
 
-	// Next, filter all txs from mem.txs that are in blockTxsMap
+	// Now we filter all txs from mem.txs that are in blockTxsMap,
-	txs := []types.Tx{}
+	// and ExecTx on what remains. Only valid txs are kept.
-	for _, tx := range mem.txs {
+	// We track the ranges of txs included in the block and invalidated by it
+	// so we can tell peer routines
+	var ri = ResetInfo{Height: block.Height}
+	var validTxs []types.Tx
+	includedStart, invalidStart := -1, -1
+	for i, tx := range mem.txs {
 		txID := types.TxID(state.ChainID, tx)
 		if _, ok := blockTxsMap[string(txID)]; ok {
+			startRange(&includedStart, i)           // start counting included txs
+			endRange(&invalidStart, i, &ri.Invalid) // stop counting invalid txs
 			log.Info("Filter out, already committed", "tx", tx, "txID", txID)
-			continue
 		} else {
-			log.Info("Filter in, still new", "tx", tx, "txID", txID)
+			endRange(&includedStart, i, &ri.Included) // stop counting included txs
-			txs = append(txs, tx)
+			err := sm.ExecTx(mem.cache, tx, false, nil)
-		}
+			if err != nil {
-	}
+				startRange(&invalidStart, i) // start counting invalid txs
-
+				log.Info("Filter out, no longer valid", "tx", tx, "error", err)
-	// Next, filter all txs that aren't valid given new state.
+			} else {
-	validTxs := []types.Tx{}
+				endRange(&invalidStart, i, &ri.Invalid) // stop counting invalid txs
-	for _, tx := range txs {
+				log.Info("Filter in, new, valid", "tx", tx, "txID", txID)
-		err := sm.ExecTx(mem.cache, tx, false, nil)
+				validTxs = append(validTxs, tx)
-		if err == nil {
+			}
-			log.Info("Filter in, valid", "tx", tx)
-			validTxs = append(validTxs, tx)
-		} else {
-			// tx is no longer valid.
-			log.Info("Filter out, no longer valid", "tx", tx, "error", err)
 		}
 	}
+	endRange(&includedStart, len(mem.txs)-1, &ri.Included) // stop counting included txs
+	endRange(&invalidStart, len(mem.txs)-1, &ri.Invalid)   // stop counting invalid txs
 
 	// We're done!
 	log.Info("New txs", "txs", validTxs, "oldTxs", mem.txs)
 	mem.txs = validTxs
+	mem.resetInfo = ri
+}
+
+func startRange(start *int, i int) {
+	if *start < 0 {
+		*start = i
+	}
+}
+
+func endRange(start *int, i int, ranger *[]Range) {
+	if *start >= 0 {
+		length := i - *start
+		*ranger = append(*ranger, Range{*start, length})
+		*start = -1
+	}
 }

mempool/reactor.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"fmt"
 	"reflect"
+	"time"
 
 	. "github.com/tendermint/tendermint/common"
 	"github.com/tendermint/tendermint/events"
@@ -14,6 +15,9 @@ import (
 
 var (
 	MempoolChannel = byte(0x30)
+
+	checkExecutedTxsMilliseconds = 1  // check for new mempool txs to send to peer
+	txsToSendPerCheck            = 64 // send up to this many txs from the mempool per check
 )
 
 // MempoolReactor handles mempool tx broadcasting amongst peers.
@@ -44,11 +48,14 @@ func (memR *MempoolReactor) GetChannels() []*p2p.ChannelDescriptor {
 }
 
 // Implements Reactor
-func (pexR *MempoolReactor) AddPeer(peer *p2p.Peer) {
+func (memR *MempoolReactor) AddPeer(peer *p2p.Peer) {
+	// Each peer gets a go routine on which we broadcast transactions in the same order we applied them to our state.
+	go memR.broadcastTxRoutine(peer)
 }
 
 // Implements Reactor
-func (pexR *MempoolReactor) RemovePeer(peer *p2p.Peer, reason interface{}) {
+func (memR *MempoolReactor) RemovePeer(peer *p2p.Peer, reason interface{}) {
+	// broadcast routine checks if peer is gone and returns
 }
 
 // Implements Reactor
@@ -70,29 +77,116 @@ func (memR *MempoolReactor) Receive(chID byte, src *p2p.Peer, msgBytes []byte) {
 		} else {
 			log.Info("Added valid tx", "tx", msg.Tx)
 		}
-		// Share tx.
+		// broadcasting happens from go routines per peer
-		// We use a simple shotgun approach for now.
-		// TODO: improve efficiency
-		for _, peer := range memR.Switch.Peers().List() {
-			if peer.Key == src.Key {
-				continue
-			}
-			peer.TrySend(MempoolChannel, msg)
-		}
-
 	default:
 		log.Warn(Fmt("Unknown message type %v", reflect.TypeOf(msg)))
 	}
 }
 
+// Just an alias for AddTx since broadcasting happens in peer routines
 func (memR *MempoolReactor) BroadcastTx(tx types.Tx) error {
-	err := memR.Mempool.AddTx(tx)
+	return memR.Mempool.AddTx(tx)
-	if err != nil {
+}
-		return err
+
+type PeerState interface {
+	GetHeight() int
+}
+
+// send new mempool txs to peer, strictly in order we applied them to our state.
+// new blocks take chunks out of the mempool, but we've already sent some txs to the peer.
+// so we wait to hear that the peer has progressed to the new height, and then continue sending txs from where we left off
+func (memR *MempoolReactor) broadcastTxRoutine(peer *p2p.Peer) {
+	newBlockChan := make(chan ResetInfo)
+	memR.evsw.(*events.EventSwitch).AddListenerForEvent("broadcastRoutine:"+peer.Key, types.EventStringNewBlock(), func(data types.EventData) {
+		// no lock needed because consensus is blocking on this
+		// and the mempool is reset before this event fires
+		newBlockChan <- memR.Mempool.resetInfo
+	})
+	timer := time.NewTicker(time.Millisecond * time.Duration(checkExecutedTxsMilliseconds))
+	currentHeight := memR.Mempool.state.LastBlockHeight
+	var nTxs, txsSent int
+	var txs []types.Tx
+	for {
+		select {
+		case <-timer.C:
+			if !peer.IsRunning() {
+				return
+			}
+
+			// make sure the peer is up to date
+			peerState := peer.Data.Get(types.PeerStateKey).(PeerState)
+			if peerState.GetHeight() < currentHeight {
+				continue
+			}
+
+			// check the mempool for new transactions
+			nTxs, txs = memR.getNewTxs(txsSent, currentHeight)
+
+			theseTxsSent := 0
+			start := time.Now()
+		TX_LOOP:
+			for _, tx := range txs {
+				// send tx to peer.
+				msg := &TxMessage{Tx: tx}
+				success := peer.Send(MempoolChannel, msg)
+				if !success {
+					break TX_LOOP
+				} else {
+					theseTxsSent += 1
+				}
+			}
+			if theseTxsSent > 0 {
+				txsSent += theseTxsSent
+				log.Warn("Sent txs to peer", "ntxs", theseTxsSent, "took", time.Since(start), "total_sent", txsSent, "total_exec", nTxs)
+			}
+
+		case ri := <-newBlockChan:
+			currentHeight = ri.Height
+
+			// find out how many txs below what we've sent were included in a block and how many became invalid
+			included := tallyRangesUpTo(ri.Included, txsSent)
+			invalidated := tallyRangesUpTo(ri.Invalid, txsSent)
+
+			txsSent -= included + invalidated
+		}
 	}
-	msg := &TxMessage{Tx: tx}
+}
-	memR.Switch.Broadcast(MempoolChannel, msg)
+
-	return nil
+// fetch new txs from the mempool
+func (memR *MempoolReactor) getNewTxs(txsSent, height int) (nTxs int, txs []types.Tx) {
+	memR.Mempool.mtx.Lock()
+	defer memR.Mempool.mtx.Unlock()
+
+	// if the mempool got ahead of us just return empty txs
+	if memR.Mempool.state.LastBlockHeight != height {
+		return
+	}
+
+	nTxs = len(memR.Mempool.txs)
+	if txsSent < nTxs {
+		if nTxs > txsSent+txsToSendPerCheck {
+			txs = memR.Mempool.txs[txsSent : txsSent+txsToSendPerCheck]
+		} else {
+			txs = memR.Mempool.txs[txsSent:]
+		}
+	}
+	return
+}
+
+// return the size of ranges less than upTo
+func tallyRangesUpTo(ranger []Range, upTo int) int {
+	totalUpTo := 0
+	for _, r := range ranger {
+		if r.Start >= upTo {
+			break
+		}
+		if r.Start+r.Length-1 > upTo {
+			totalUpTo += upTo - r.Start - 1
+			break
+		}
+		totalUpTo += r.Length
+	}
+	return totalUpTo
 }
 
 // implements events.Eventable

rpc/core/consensus.go

@@ -31,7 +31,7 @@ func DumpConsensusState() (*ctypes.ResultDumpConsensusState, error) {
 	peerRoundStates := []string{}
 	for _, peer := range p2pSwitch.Peers().List() {
 		// TODO: clean this up?
-		peerState := peer.Data.Get(cm.PeerStateKey).(*cm.PeerState)
+		peerState := peer.Data.Get(types.PeerStateKey).(*cm.PeerState)
 		peerRoundState := peerState.GetRoundState()
 		peerRoundStateStr := peer.Key + ":" + string(wire.JSONBytes(peerRoundState))
 		peerRoundStates = append(peerRoundStates, peerRoundStateStr)

rpc/core/mempool.go

@@ -30,5 +30,6 @@ func BroadcastTx(tx types.Tx) (*ctypes.ResultBroadcastTx, error) {
 }
 
 func ListUnconfirmedTxs() (*ctypes.ResultListUnconfirmedTxs, error) {
-	return &ctypes.ResultListUnconfirmedTxs{mempoolReactor.Mempool.GetProposalTxs()}, nil
+	txs := mempoolReactor.Mempool.GetProposalTxs()
+	return &ctypes.ResultListUnconfirmedTxs{len(txs), txs}, nil
 }

rpc/core/types/responses.go

@@ -98,6 +98,7 @@ type Receipt struct {
 }
 
 type ResultListUnconfirmedTxs struct {
+	N   int        `json:"n_txs"`
 	Txs []types.Tx `json:"txs"`
 }