In this commit:
1. Request unknown MPT nodes from peers. Note, that StateSync module itself
shouldn't be responsible for nodes requests, that's a server duty.
2. Do not request the same node twice, check if it is in storage
already. If so, then the only thing remaining is to update refcounter.
Refactor code and be fine with sending to just 2/3 of proper peers. Previously
it was an edge case, but it can be a normal thing to do also as broadcasting
to everyone is obviously too expensive and excessive (hi, #608).
Baseline (four node, 10 workers):
RPS 8180.760 8137.822 7858.358 7820.011 8051.076 ≈ 8010 ± 2.04%
TPS 7819.831 7521.172 7519.023 7242.965 7426.000 ≈ 7506 ± 2.78%
CPU % 41.983 38.775 40.606 39.375 35.537 ≈ 39.3 ± 6.15%
Mem MB 2947.189 2743.658 2896.688 2813.276 2863.108 ≈ 2853 ± 2.74%
Patched:
RPS 9714.567 9676.102 9358.609 9371.408 9301.372 ≈ 9484 ± 2.05% ↑ 18.40%
TPS 8809.796 8796.854 8534.754 8661.158 8426.162 ≈ 8646 ± 1.92% ↑ 15.19%
CPU % 44.980 45.018 33.640 29.645 43.830 ≈ 39.4 ± 18.41% ↑ 0.25%
Mem MB 2989.078 2976.577 2306.185 2351.929 2910.479 ≈ 2707 ± 12.80% ↓ 5.12%
There is a nuance with this patch however. While typically it works the way
outlined above, sometimes it works like this:
RPS ≈ 6734.368
TPS ≈ 6299.332
CPU ≈ 25.552%
Mem ≈ 2706.046MB
And that's because the log looks like this:
DeltaTime, TransactionsCount, TPS
5014, 44212, 8817.710
5163, 49690, 9624.249
5166, 49523, 9586.334
5189, 49693, 9576.604
5198, 49339, 9491.920
5147, 49559, 9628.716
5192, 49680, 9568.567
5163, 49750, 9635.871
5183, 49189, 9490.450
5159, 49653, 9624.540
5167, 47945, 9279.079
5179, 2051, 396.022
5015, 4, 0.798
5004, 0, 0.000
5003, 0, 0.000
5003, 0, 0.000
5003, 0, 0.000
5003, 0, 0.000
5004, 0, 0.000
5003, 2925, 584.649
5040, 49099, 9741.865
5161, 49718, 9633.404
5170, 49228, 9521.857
5179, 49773, 9610.543
5167, 47253, 9145.152
5202, 49788, 9570.934
5177, 47704, 9214.603
5209, 46610, 8947.975
5249, 49156, 9364.831
5163, 18284, 3541.352
5072, 174, 34.306
On a network with 4 CNs and 1 RPC node there is 1/256 probability that a block
won't be broadcasted to RPC node, so it won't see it until ping timeout kicks
in. While it doesn't see a block it can't accept new incoming transactions so
the bench gets stuck basically. To me that's an acceptable trade-off because
normal networks are much larger than that and the effect of this patch is way
more important there, but still that's what we have and we need to take into
account.
send() can return errStateMismatch, errGone and errBusy. errGone means the
peer is dead and it won't ever be active again, it doesn't make sense retrying
sends to it. errStateMismatch is technically "not yet ready", but we can't
wait for it either, no one knows how much will it take to complete
handshake. So only errBusy means we can retry.
So keep track of dead peers and adjust tries counting appropriately.
It doesn't change much, we can't magically get more valid peers and if some
die while we're iterating we'd detect that by an error returned from send().
When transaction spreads through the network many nodes are likely to get it
in roughly the same time. They will rebroadcast it also in roughly the same
time. As we have a number of peers it's quite likely that we'd get an Inv with
the same transaction from multiple peers simultaneously. We will ask them for
this transaction (independently!) and again we're likely to get it in roughly
the same time. So we can easily end up with multiple threads processing the
same transaction. Only one will succeed, but we can actually easily avoid
doing it in the first place saving some CPU cycles for other things.
Notice that we can't do it _before_ receiving a transaction because nothing
guarantees that the peer will respond to our transaction request, so
communication overhead is unavoidable at the moment, but saving on processing
already gives quite interesting results.
Baseline, four nodes with 10 workers:
RPS 7176.784 7014.511 6139.663 7191.280 7080.852 ≈ 6921 ± 5.72%
TPS 6945.409 6562.756 5927.050 6681.187 6821.794 ≈ 6588 ± 5.38%
CPU % 44.400 43.842 40.418 49.211 49.370 ≈ 45.4 ± 7.53%
Mem MB 2693.414 2640.602 2472.007 2731.482 2707.879 ≈ 2649 ± 3.53%
Patched:
RPS ≈ 7791.675 7996.559 7834.504 7746.705 7891.614 ≈ 7852 ± 1.10% ↑ 13.45%
TPS ≈ 7241.497 7711.765 7520.211 7425.890 7334.443 ≈ 7447 ± 2.17% ↑ 13.04%
CPU % 29.853 39.936 39.945 36.371 39.999 ≈ 37.2 ± 10.57% ↓ 18.06%
Mem MB 2749.635 2791.609 2828.610 2910.431 2863.344 ≈ 2829 ± 1.97% ↑ 6.80%
Most of the time on healthy network we see new transactions appearing that are
not present in the mempool. Once they get into mempool we don't ask for them
again when some other peer sends an Inv with them. Then these transactions are
usually added into block, removed from mempool and no one actually sends them
again to us. Some stale nodes can do that, but it's not very likely to
happen.
At the receiving end at the same time it's quite expensive to do full chain
HasTransaction() query, so if we can avoid doing that it's always good. Here
it technically allows resending old transaction that will be re-requested and
an attempt to add it to mempool will be made. But it'll inevitably fail
because the same HasTransaction() check is done there too. One can try to
maliciously flood the node with stale transactions but it doesn't differ from
flooding it with any other invalid transactions, so there is no new attack
vector added.
Baseline, 4 nodes with 10 workers:
RPS 6902.296 6465.662 6856.044 6785.515 6157.024 ≈ 6633 ± 4.26%
TPS 6468.431 6218.867 6610.565 6288.596 5790.556 ≈ 6275 ± 4.44%
CPU % 50.231 42.925 49.481 48.396 42.662 ≈ 46.7 ± 7.01%
Mem MB 2856.841 2684.103 2756.195 2733.485 2422.787 ≈ 2691 ± 5.40%
Patched:
RPS 7176.784 7014.511 6139.663 7191.280 7080.852 ≈ 6921 ± 5.72% ↑ 4.34%
TPS 6945.409 6562.756 5927.050 6681.187 6821.794 ≈ 6588 ± 5.38% ↑ 4.99%
CPU % 44.400 43.842 40.418 49.211 49.370 ≈ 45.4 ± 7.53% ↓ 2.78%
Mem MB 2693.414 2640.602 2472.007 2731.482 2707.879 ≈ 2649 ± 3.53% ↓ 1.56%
There was a deadlock while trying to finalize transaction during
PostBlock:
1) (*Notary).PostBlock is called under the blockchain lock
2) (*Notary).onTransaction is called inside the PostBlock
3) (*Notary).onTransaction needs to RLock the blockchain to add
completed transaction to the memory pool (and the blockchain is Lock'ed
by this moment)
The problem is fixed by using notifications subsistem, because it's not
required to call (*Notary).PostBlock under the blockchain lock.
And stop dropping connections if we're to receive them. Proper handling is
subject of #1701, but we need at least some connection-level stability for
now.
Node receiving extensible payload from the future is confused and drops
connection. Note that this can still happen if the node is to loose its
synchrony.
Calling `IsInSync()` is quite expensive, so we stop doing that once synchrony
is reached (hence bool flag).
1. Initialization is performed via `Blockchain` methods.
2. Native Oracle contract updates list of oracle nodes
and in-fly requests in `PostPersist`.
3. RPC uses Oracle module directly.
It could be the case that checks are performed simultaneosly and
peers connections goes down from 2 to 0. We must take such case into
account and register address as good in discovery.
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.
If the node is to start with seeds unavailable it will try connecting to each
of them three times, blacklist them and then sit forever waiting for
something. It's not a good behavior, it should always try connecting to seeds
if nothing else works.
GetBlockByIndex handler starts sending blocks right from the start index and
if that index is s.chain.BlockHeight() then we're requesting and receiving a
block we already have.
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.
It returned an error in case if block wasn't found (it might be when our
chain is lower). Fixed. It also should return all requested blocks, not
the first one.
We can't lock them (or there will be a deadlock), but we need to fix this:
fatal error: concurrent map iteration and map write
goroutine 1 [running]:
runtime.throw(0xdec086, 0x26)
/usr/lib64/go/1.12/src/runtime/panic.go:617 +0x72 fp=0xc02fec2bf8 sp=0xc02fec2bc8 pc=0x42d932
runtime.mapiternext(0xc02fec2d40)
/usr/lib64/go/1.12/src/runtime/map.go:860 +0x597 fp=0xc02fec2c80 sp=0xc02fec2bf8 pc=0x40efe7
github.com/nspcc-dev/neo-go/pkg/network.(*Server).Shutdown(0xc0000fc160)
/home/rik/dev/neo-go2/pkg/network/server.go:194 +0x238 fp=0xc02fec2db0 sp=0xc02fec2c80 pc=0xa89da8
github.com/nspcc-dev/neo-go/cli/server.startServer(0xc0000fcc60, 0x0, 0x0)
/home/rik/dev/neo-go2/cli/server/server.go:399 +0x7a9 fp=0xc02fec3820 sp=0xc02fec2db0 pc=0xae2079
...
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.
Get new blocks directly from the Blockchain. It may lead to some duplications
(as we'll also receive our own blocks), but at the same time it's more
correct, because technically we can also get blocks via other means besides
network server like RPC (submitblock call). And it simplifies network server
at the same time.
Close transport and disconnect peers right in the Shutdown(), so that no new
connections would be accepted and so that all the peers would be disconnected
correctly (avoiding the same deadlock as in e2116e4c3f).
Implement mempool and consensus block creation policies, almost the same as
SimplePolicy plugin for C# node provides with two caveats:
* HighPriorityTxType is not configured and hardcoded to ClaimType
* BlockedAccounts are not supported
Other than that it allows us to run successfuly as testnet CN, previously our
proposals were rejected because we were proposing blocks with oversized
transactions (that are rejected by PoolTx() now).
Mainnet and testnet configuration files are updated accordingly, but privnet
is left as is with no limits.
Configuration is currently attached to the Blockchain and so is the code that
does policying, it may be moved somewhere in the future, but it works for
now.
We can only add one block of the given height and we have two competing
goroutines to do that --- consensus and block queue. Whomever adds the block
first shouldn't trigger an error in another one.
Fix block relaying for blocks added via the block queue also, previously one
consensus-generated blocks were broadcasted.
Eliminate races between tx checks and adding them to the mempool, ensure the
chain doesn't change while we're working with the new tx. Ensure only one
block addition attempt could be in progress.
It can lead to some goroutine explosion, but supposedly it's better than
stalling other processing and eventually all of these goroutines should finish
their sends. Note that this doesn't change the behavior for RPC-relayed
transactions that are still waiting for the broadcast to finish ensuring
proper transaction distribution before returning the result to the client.
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).
They have the opposite order, height first and nonce second. It was done wrong
in 4e6ed902 and never fixed since. Fixes sending wrong peer state leading to
useless getheaders messages (and disconnects when the other side is lagging
behind).
We can have more than one connection attempt in progress and not yet completed
the handshake, so if there is a Version already received we should look it.
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.
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)
It wasn't actually requesting transactions but rather sending an inventory
message telling everyone that we have them which is completely wrong and
easily leads to ChangeView that could be avoided.
Only request headers from the other peer if his height is bigger than
ours. Otherwise we routinely ask 0-height newcomers for some random headers
that they know nothing about.
This one is essential for the consensus nodes as otherwise they won't give out
the blocks they generate making their generation almost useless. It also makes
our networking part more complete.
We have a race between reader and writer goroutines for the same connection
that leads to handshake failures when reader is faster to read the incoming
version (and try to reply to it) than writer is to write our own Version:
WARN[0000] peer disconnected addr="172.200.0.4:20334" peerCount=5 reason="invalid handshake: tried to send VersionAck, but didn't send Version yet
Fix it by moving Version sending before the reader loop starts.
Our node didn't respect the MaxPeers setting, fix it with a drop of random
connection when this limit is reached (to give a chance for newcomers to
communicate), but also introduce AttemptConnPeers setting to tune the number
of attempted connections.
This also raises the default MaxPeers for testnet/mainnet to 100, because
neo-go nodes love making friends.
