This is not exactly the protocol-level batching as was tried in #1770 and
proposed by neo-project/neo#2365, but it's a TCP-level change in that we now
Write() a set of messages and given that Go sets up TCP sockets with
TCP_NODELAY by default this is a substantial change, we have less packets
generated with the same amount of data. It doesn't change anything on properly
connected networks, but the ones with delays benefit from it a lot.
This also improves queueing because we no longer generate 32 messages to
deliver on transaction's GetData, it's just one stream of bytes with 32
messages inside.
Do the same with GetBlocksByIndex, we can have a lot of messages there too.
But don't forget about potential peer DoS attacks, if a peer is to request a
lot of big blocks we need to flush them before we process the whole set.
Drop EnqueueP2PPacket, replace EnqueueHPPacket with EnqueueHPMessage. We use
Enqueue* when we have a specific per-peer message, it makes zero sense
duplicating serialization code for it (unlike Broadcast*).
Follow the general rules of broadcasts, even though it's somewhat different
from Inv, we just want to get some reply from our neighbors to see if we're
behind. We don't strictly need all neighbors for it.
We have a number of queues for different purposes:
* regular broadcast queue
* direct p2p queue
* high-priority queue
And two basic egress scenarios:
* direct p2p messages (replies to requests in Server's handle* methods)
* broadcasted messages
Low priority broadcasted messages:
* transaction inventories
* block inventories
* notary inventories
* non-consensus extensibles
High-priority broadcasted messages:
* consensus extensibles
* getdata transaction requests from consensus process
* getaddr requests
P2P messages are a bit more complicated, most of the time they use p2p queue,
but extensible message requests/replies use HP queue.
Server's handle* code is run from Peer's handleIncoming, every peer has this
thread that handles incoming messages. When working with the peer it's
important to reply to requests and blocking this thread until we send (queue)
a reply is fine, if the peer is slow we just won't get anything new from
it. The queue used is irrelevant wrt this issue.
Broadcasted messages are radically different, we want them to be delivered to
many peers, but we don't care about specific ones. If it's delivered to 2/3 of
the peers we're fine, if it's delivered to more of them --- it's not an
issue. But doing this fairly is not an easy thing, current code tries performing
unblocked sends and if this doesn't yield enough results it then blocks (but
has a timeout, we can't wait indefinitely). But it does so in sequential
manner, once the peer is chosen the code will wait for it (and only it) until
timeout happens.
What can be done instead is an attempt to push the message to all of the peers
simultaneously (or close to that). If they all deliver --- OK, if some block
and wait then we can wait until _any_ of them pushes the message through (or
global timeout happens, we still can't wait forever). If we have enough
deliveries then we can cancel pending ones and it's again not an error if
these canceled threads still do their job.
This makes the system more dynamic and adds some substantial processing
overhead, but it's a networking code, any of this overhead is much lower than
the actual packet delivery time. It also allows to spread the load more
fairly, if there is any spare queue it'll get the packet and release the
broadcaster. On the next broadcast iteration another peer is more likely to be
chosen just because it didn't get a message previously (and had some time to
deliver already queued messages).
It works perfectly in tests, with optimal networking conditions we have much
better block times and TPS increases by 5-25%% depending on the scenario.
I'd go as far as to say that it fixes the original problem of #2678, because
in this particular scenario we have empty queues in ~100% of the cases and
this new logic will likely lead to 100% fan out in this case (cancelation just
won't happen fast enough). But when the load grows and there is some waiting
in the queue it will optimize out the slowest links.
Peers can be slow, very slow, slow enough to affect node's regular
operation. We can't wait for them indefinitely, there has to be a timeout for
send operations.
This patch uses TimePerBlock as a reference for its timeout. It's relatively
big and it doesn't affect tests much, 4+1 scenarios tend to perform a little
worse with while 7+2 scenarios work a little better. The difference is in some
percents, but all of these tests easily have 10-15% variations from run to
run.
It's an important step in making our gossip better because we can't have any
behavior where neighbors directly block the node forever, refs. #2678 and
Network communication takes time. Handling some messages (like transaction)
also takes time. We can share this time by making handler a separate
goroutine. So while message is being handled receiver can already get and
parse the next one.
It doesn't improve metrics a lot, but still I think it makes sense and in some
scenarios this can be more beneficial than this.
e41fc2fd1b, 4 nodes, 10 workers
RPS 6732.979 6396.160 6759.624 6246.398 6589.841 ≈ 6545 ± 3.02%
TPS 6491.062 5984.190 6275.652 5867.477 6360.797 ≈ 6196 ± 3.77%
CPU % 42.053 43.515 44.768 40.344 44.112 ≈ 43.0 ± 3.69%
Mem MB 2564.130 2744.236 2636.267 2589.505 2765.926 ≈ 2660 ± 3.06%
Patched:
RPS 6902.296 6465.662 6856.044 6785.515 6157.024 ≈ 6633 ± 4.26% ↑ 1.34%
TPS 6468.431 6218.867 6610.565 6288.596 5790.556 ≈ 6275 ± 4.44% ↑ 1.28%
CPU % 50.231 42.925 49.481 48.396 42.662 ≈ 46.7 ± 7.01% ↑ 8.60%
Mem MB 2856.841 2684.103 2756.195 2733.485 2422.787 ≈ 2691 ± 5.40% ↑ 1.17%
Right now a single slow peer can slow down whole network.
Do broadcast in 2 parts:
1. Perform non-blocking send to all peers if possible.
2. Perform blocking sends until message is sent to 2/3 of good peers.
Turns out, C# node no longer broadcasts an Inv when it's creating a block,
instead it sends a ping and if we're not paying attention to the height
specified there we're technically missing a new block. Of course we'll get it
later after ping timer expiration and regular ping/pong sequence, but that's
delaying it for no good reason.
Closes#1192
1. We now have CMDGetBlockByIndex, so there's no need to request headers
first when we can just ask for blocks.
2. We don't ask for headers (i.e. we don't send CMDGetHeaders),
consequently, we shouldn't react on CMDHeaders.
3. But we still keep on reacting on CMDGetHeaders command as
there could be a node which needs headers.
We make it explicit in the appropriate Block/Transaction structures, not via a
singleton as C# node does. I think this approach has a bit more potential and
allows better packages reuse for different purposes.
We can still lock the (*Server).run with dead peers:
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: goroutine 40 [select, 871 minutes]:
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).putPacketIntoQueue(0xc030ab5320, 0xc02f251f20, 0xc00af0dcc0, 0x18, 0x40, 0x100000000000000, 0xffffffffffffffff)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:82 +0xf4
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).EnqueueHPPacket(0xc030ab5320, 0xc00af0dcc0, 0x18, 0x40, 0x1367240, 0xc03090ef98)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:124 +0x52
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*Server).iteratePeersWithSendMsg(0xc0000ca000, 0xc00af35800, 0xcb2a58, 0x0)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:720 +0x12a
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*Server).broadcastHPMessage(...)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:731
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*Server).run(0xc0000ca000)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:203 +0xee4
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*Server).Start(0xc0000ca000, 0xc000072ba0)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:173 +0x2ec
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: created by github.com/CityOfZion/neo-go/cli/server.startServer
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/cli/server/server.go:331 +0x476
...
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: goroutine 2199 [chan send, 870 minutes]:
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).Disconnect.func1()
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:366 +0x85
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: sync.(*Once).Do(0xc030ab403c, 0xc02f262788)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/usr/local/go/src/sync/once.go:44 +0xb3
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).Disconnect(0xc030ab4000, 0xd92440, 0xc000065a00)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:365 +0x6d
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).SendPing.func1()
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:394 +0x42
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: created by time.goFunc
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/usr/local/go/src/time/sleep.go:169 +0x44
...
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: goroutine 3448 [chan send, 854 minutes]:
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).handleConn(0xc01ed203f0)
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:143 +0x6c
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: created by github.com/CityOfZion/neo-go/pkg/network.(*TCPTransport).Accept
Feb 13 16:14:50 neo-go-node-2 neo-go[9448]: #011/go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_transport.go:62 +0x44c
...
The problem is that the select in putPacketIntoQueue() only works the way it
was intended to after the `close(p.done)`, but that happens only after
successful unregistration request send. Thus, do disconnects the other way
around, first unblock queueing and exit goroutines, then destroy the
connection (if it wasn't previously destroyed) and only after that signal to
the Server.
We can leak sending goroutines and stall broadcasts because of already gone
peers that happened to be cached by some s.Peers() user (more than 800 of
these can be seen in nodoka log along with (*Server).run blocking on
CMDGetAddr send):
Feb 10 16:35:15 nodoka neo-go[1563]: goroutine 41 [chan send, 3320 minutes]:
Feb 10 16:35:15 nodoka neo-go[1563]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).putPacketIntoQueue(...)
Feb 10 16:35:15 nodoka neo-go[1563]: /go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:81
Feb 10 16:35:15 nodoka neo-go[1563]: github.com/CityOfZion/neo-go/pkg/network.(*TCPPeer).EnqueueHPPacket(0xc0083d57a0, 0xc017206100, 0x18, 0x40, 0x136a240, 0xc018ef9720)
Feb 10 16:35:15 nodoka neo-go[1563]: /go/src/github.com/CityOfZion/neo-go/pkg/network/tcp_peer.go:119 +0x98
Feb 10 16:35:15 nodoka neo-go[1563]: github.com/CityOfZion/neo-go/pkg/network.(*Server).iteratePeersWithSendMsg(0xc0000ca000, 0xc0001848a0, 0xcb4550, 0x0)
Feb 10 16:35:15 nodoka neo-go[1563]: /go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:720 +0x12a
Feb 10 16:35:15 nodoka neo-go[1563]: github.com/CityOfZion/neo-go/pkg/network.(*Server).broadcastHPMessage(...)
Feb 10 16:35:15 nodoka neo-go[1563]: /go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:731
Feb 10 16:35:15 nodoka neo-go[1563]: github.com/CityOfZion/neo-go/pkg/network.(*Server).run(0xc0000ca000)
Feb 10 16:35:15 nodoka neo-go[1563]: /go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:203 +0xee4
Feb 10 16:35:15 nodoka neo-go[1563]: github.com/CityOfZion/neo-go/pkg/network.(*Server).Start(0xc0000ca000, 0xc000072c60)
Feb 10 16:35:15 nodoka neo-go[1563]: /go/src/github.com/CityOfZion/neo-go/pkg/network/server.go:173 +0x2ec
Feb 10 16:35:15 nodoka neo-go[1563]: created by github.com/CityOfZion/neo-go/cli/server.startServer
Feb 10 16:35:15 nodoka neo-go[1563]: /go/src/github.com/CityOfZion/neo-go/cli/server/server.go:331 +0x476
If we have already got Version message, we don't need the rest of handshake to
complete before being able to properly answer the PeerAddr() requests. Fixes
some duplicate connections between machines.
This one is designed to give more priority to direct nodes communication, that
is that their messaging would have more priority than generic broadcasts. It
should improve consensus process under TX pressure and allow to handle
pings in time (preventing disconnects).
Our node was too pingy because of wrong timer setups (that divided timeout
Duration by time.Second), it also was wrong in its time calculations (using
UTC time to calculate intervals). At the same time missing block is a
server-wide problem, so it's better solved with server-wide protocol loop.
1) Make timeout a timeout, don't do magic ping counts.
2) Drop additional timer from the main peer's protocol loop, create it
dynamically and make it disconnect the peer.
3) Don't expose the ping counter to the outside, handle more logic inside the
Peer.
Relates to #430.
We don't and we won't have synchronized clocks in the network so the only
timestamp that we can compare our local time with is the one made
ourselves. What this ping mechanism is used for is to recover from missing the
block broadcast, thus it's appropriate for it to trigger after X seconds of
the local time since the last block received.
Relates to #430.
In reality it will never be true exactly in the case where we want this ping
mechanism to work --- when the node failed to get a block from the net. It
won't get the header either and thus its block height will be equal to header
height. The only moment when this condition is met is when the node does
initial synchronization and this synchronization works just fine without any
pings.
Relates to #430.
Two queues for high-priority and ordinary messages. Fixes#590. These queues
are deliberately made small to avoid buffer bloat problem, there is gonna be
another queueing layer above them to compensate for that. The queues are
designed to be synchronous in enqueueing, async capabilities are to be added
layer above later.
add pingInterval same as used in ref C# implementation with the same logic
add pingTimeout which is used to check whether pong received. If not -- drop the peer.
add pingLimit which is hardcoded to 4 in TCPPeer. It's limit for unsuccessful ping/pong calls (where pong wasn't received in pingTimeout interval)
Right now message can be written in several Write's so
concurrent calls of writeMsg() can in theory interleave.
This commit fixes it.
Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
Otherwise the node might crash in `startProtocol` because of missing Version
field in the peer. And it also keeps the sequence correct, Version MUST be
sent first and ACKs can only follow it.
This allows to start handshaking from both client and server (mainnet/testnet
nodes were seen to not care about string ordering for it), but still maintains
some sane checks in the process. It also makes functions thread-safe because
we have two goroutines servicing read and write side of the Peer connection,
so they can clash on access to the struct fields.
Add a test for it also.
