It makes sense in general (further narrowing down the time window when
transactions are processed by consensus thread) and it improves block times a
little too, especially in the 7+2 scenario.
Related to #2744.
Until the consensus process starts for a new block and until it really needs
some transactions we can spare some cycles by not delivering transactions to
it. In tests this doesn't affect TPS, but makes block delays a bit more
stable. Related to #2744, I think it also may cause timeouts during
transaction processing (waiting on the consensus process channel while it does
something dBFT-related).
When the network is big enough, MinPeers may be suboptimal for good network
connectivity, but if we know the network size we can do some estimation on the
number of sufficient peers.
They can fail right in the getPeers or they can fail later when packet send
is attempted. Of course they can complete handshake in-between these events,
but most likely they won't and we'll waste more resources on this attempt. So
rule out bad peers immediately.
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
Blockchain's notificationDispatcher sends events to channels and these
channels must be read from. Unfortunately, regular service shutdown procedure
does unsubscription first (outside of the read loop) and only then drains the
channel. While it waits for unsubscription request to be accepted
notificationDispatcher can try pushing more data into the same channel which
will lead to a deadlock. Reading in the same method solves this, any number of
events can be pushed until unsub channel accepts the data.
Unsubscribe and drain first, then return from the Shutdown method. It's
important wrt to subsequent chain shutdown process (normally it's closed right
after the network server).
Consensus can require conflicting transactions and it can require more
transactions than mempool can fit, all of this should work. Transactions will
be checked anyway using its secondary mempool. See the scenario from #668.
Notice that it makes the node accept Extensible payloads with any category
which is the same way C# node works. We're trusting Extensible senders,
improper payloads are harmless until they DoS the network, but we have some
protections against that too (and spamming with proper category doesn't differ
a lot).
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.