go-pulse/les/distributor.go
Felföldi Zsolt c2003ed63b les, les/flowcontrol: improved request serving and flow control (#18230)
This change

- implements concurrent LES request serving even for a single peer.
- replaces the request cost estimation method with a cost table based on
  benchmarks which gives much more consistent results. Until now the
  allowed number of light peers was just a guess which probably contributed
  a lot to the fluctuating quality of available service. Everything related
  to request cost is implemented in a single object, the 'cost tracker'. It
  uses a fixed cost table with a global 'correction factor'. Benchmark code
  is included and can be run at any time to adapt costs to low-level
  implementation changes.
- reimplements flowcontrol.ClientManager in a cleaner and more efficient
  way, with added capabilities: There is now control over bandwidth, which
  allows using the flow control parameters for client prioritization.
  Target utilization over 100 percent is now supported to model concurrent
  request processing. Total serving bandwidth is reduced during block
  processing to prevent database contention.
- implements an RPC API for the LES servers allowing server operators to
  assign priority bandwidth to certain clients and change prioritized
  status even while the client is connected. The new API is meant for
  cases where server operators charge for LES using an off-protocol mechanism.
- adds a unit test for the new client manager.
- adds an end-to-end test using the network simulator that tests bandwidth
  control functions through the new API.
2019-02-26 12:32:48 +01:00

286 lines
8.0 KiB
Go

// Copyright 2017 The go-ethereum Authors
// This file is part of the go-ethereum library.
//
// The go-ethereum library is free software: you can redistribute it and/or modify
// it under the terms of the GNU Lesser General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// The go-ethereum library is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
package les
import (
"container/list"
"sync"
"time"
"github.com/ethereum/go-ethereum/common/mclock"
)
// requestDistributor implements a mechanism that distributes requests to
// suitable peers, obeying flow control rules and prioritizing them in creation
// order (even when a resend is necessary).
type requestDistributor struct {
clock mclock.Clock
reqQueue *list.List
lastReqOrder uint64
peers map[distPeer]struct{}
peerLock sync.RWMutex
stopChn, loopChn chan struct{}
loopNextSent bool
lock sync.Mutex
}
// distPeer is an LES server peer interface for the request distributor.
// waitBefore returns either the necessary waiting time before sending a request
// with the given upper estimated cost or the estimated remaining relative buffer
// value after sending such a request (in which case the request can be sent
// immediately). At least one of these values is always zero.
type distPeer interface {
waitBefore(uint64) (time.Duration, float64)
canQueue() bool
queueSend(f func())
}
// distReq is the request abstraction used by the distributor. It is based on
// three callback functions:
// - getCost returns the upper estimate of the cost of sending the request to a given peer
// - canSend tells if the server peer is suitable to serve the request
// - request prepares sending the request to the given peer and returns a function that
// does the actual sending. Request order should be preserved but the callback itself should not
// block until it is sent because other peers might still be able to receive requests while
// one of them is blocking. Instead, the returned function is put in the peer's send queue.
type distReq struct {
getCost func(distPeer) uint64
canSend func(distPeer) bool
request func(distPeer) func()
reqOrder uint64
sentChn chan distPeer
element *list.Element
}
// newRequestDistributor creates a new request distributor
func newRequestDistributor(peers *peerSet, stopChn chan struct{}, clock mclock.Clock) *requestDistributor {
d := &requestDistributor{
clock: clock,
reqQueue: list.New(),
loopChn: make(chan struct{}, 2),
stopChn: stopChn,
peers: make(map[distPeer]struct{}),
}
if peers != nil {
peers.notify(d)
}
go d.loop()
return d
}
// registerPeer implements peerSetNotify
func (d *requestDistributor) registerPeer(p *peer) {
d.peerLock.Lock()
d.peers[p] = struct{}{}
d.peerLock.Unlock()
}
// unregisterPeer implements peerSetNotify
func (d *requestDistributor) unregisterPeer(p *peer) {
d.peerLock.Lock()
delete(d.peers, p)
d.peerLock.Unlock()
}
// registerTestPeer adds a new test peer
func (d *requestDistributor) registerTestPeer(p distPeer) {
d.peerLock.Lock()
d.peers[p] = struct{}{}
d.peerLock.Unlock()
}
// distMaxWait is the maximum waiting time after which further necessary waiting
// times are recalculated based on new feedback from the servers
const distMaxWait = time.Millisecond * 10
// main event loop
func (d *requestDistributor) loop() {
for {
select {
case <-d.stopChn:
d.lock.Lock()
elem := d.reqQueue.Front()
for elem != nil {
req := elem.Value.(*distReq)
close(req.sentChn)
req.sentChn = nil
elem = elem.Next()
}
d.lock.Unlock()
return
case <-d.loopChn:
d.lock.Lock()
d.loopNextSent = false
loop:
for {
peer, req, wait := d.nextRequest()
if req != nil && wait == 0 {
chn := req.sentChn // save sentChn because remove sets it to nil
d.remove(req)
send := req.request(peer)
if send != nil {
peer.queueSend(send)
}
chn <- peer
close(chn)
} else {
if wait == 0 {
// no request to send and nothing to wait for; the next
// queued request will wake up the loop
break loop
}
d.loopNextSent = true // a "next" signal has been sent, do not send another one until this one has been received
if wait > distMaxWait {
// waiting times may be reduced by incoming request replies, if it is too long, recalculate it periodically
wait = distMaxWait
}
go func() {
d.clock.Sleep(wait)
d.loopChn <- struct{}{}
}()
break loop
}
}
d.lock.Unlock()
}
}
}
// selectPeerItem represents a peer to be selected for a request by weightedRandomSelect
type selectPeerItem struct {
peer distPeer
req *distReq
weight int64
}
// Weight implements wrsItem interface
func (sp selectPeerItem) Weight() int64 {
return sp.weight
}
// nextRequest returns the next possible request from any peer, along with the
// associated peer and necessary waiting time
func (d *requestDistributor) nextRequest() (distPeer, *distReq, time.Duration) {
checkedPeers := make(map[distPeer]struct{})
elem := d.reqQueue.Front()
var (
bestPeer distPeer
bestReq *distReq
bestWait time.Duration
sel *weightedRandomSelect
)
d.peerLock.RLock()
defer d.peerLock.RUnlock()
for (len(d.peers) > 0 || elem == d.reqQueue.Front()) && elem != nil {
req := elem.Value.(*distReq)
canSend := false
for peer := range d.peers {
if _, ok := checkedPeers[peer]; !ok && peer.canQueue() && req.canSend(peer) {
canSend = true
cost := req.getCost(peer)
wait, bufRemain := peer.waitBefore(cost)
if wait == 0 {
if sel == nil {
sel = newWeightedRandomSelect()
}
sel.update(selectPeerItem{peer: peer, req: req, weight: int64(bufRemain*1000000) + 1})
} else {
if bestReq == nil || wait < bestWait {
bestPeer = peer
bestReq = req
bestWait = wait
}
}
checkedPeers[peer] = struct{}{}
}
}
next := elem.Next()
if !canSend && elem == d.reqQueue.Front() {
close(req.sentChn)
d.remove(req)
}
elem = next
}
if sel != nil {
c := sel.choose().(selectPeerItem)
return c.peer, c.req, 0
}
return bestPeer, bestReq, bestWait
}
// queue adds a request to the distribution queue, returns a channel where the
// receiving peer is sent once the request has been sent (request callback returned).
// If the request is cancelled or timed out without suitable peers, the channel is
// closed without sending any peer references to it.
func (d *requestDistributor) queue(r *distReq) chan distPeer {
d.lock.Lock()
defer d.lock.Unlock()
if r.reqOrder == 0 {
d.lastReqOrder++
r.reqOrder = d.lastReqOrder
}
back := d.reqQueue.Back()
if back == nil || r.reqOrder > back.Value.(*distReq).reqOrder {
r.element = d.reqQueue.PushBack(r)
} else {
before := d.reqQueue.Front()
for before.Value.(*distReq).reqOrder < r.reqOrder {
before = before.Next()
}
r.element = d.reqQueue.InsertBefore(r, before)
}
if !d.loopNextSent {
d.loopNextSent = true
d.loopChn <- struct{}{}
}
r.sentChn = make(chan distPeer, 1)
return r.sentChn
}
// cancel removes a request from the queue if it has not been sent yet (returns
// false if it has been sent already). It is guaranteed that the callback functions
// will not be called after cancel returns.
func (d *requestDistributor) cancel(r *distReq) bool {
d.lock.Lock()
defer d.lock.Unlock()
if r.sentChn == nil {
return false
}
close(r.sentChn)
d.remove(r)
return true
}
// remove removes a request from the queue
func (d *requestDistributor) remove(r *distReq) {
r.sentChn = nil
if r.element != nil {
d.reqQueue.Remove(r.element)
r.element = nil
}
}