Satoshi Client Operation: Block Exchange
-----------------------------------------------
This article describes how blocks are exchanged between nodes.
See this article for more information on how blocks are validated:
https://en.bitcoin.it/wiki/Protocol_rules#Blocks
Upon initial connection, if the connection was not inbound[1],
or in other words, if the connection was initiated by the local node,
the version message is queued for sending immediately. When the remote node
receives the version message it replies with its own version message.[2]
When a node receives a "version" message, it may send a "getblocks" request
to the remote node if EITHER:
1. The node has never sent an initial getblocks request to any node yet.
2. Or, this is the only active node connection. Presumably the node had
zero connections prior to this connection, so maybe it was disconnected
for a long time. So, it will ask for blocks to catch up.
The getblocks message contains multiple block hashes that the requesting
node already possesses, in order to help the remote note find the latest
common block between the nodes. The list of hashes starts with the latest
block and goes back ten and then doubles in an exponential progression
until the genesis block is reached.[3] Since both nodes are hard coded
with the genesis block, they are guaranteed to a least start there.
If that block does not match for some reason, no blocks are exchanged.
-- Inventory Messages --
Note that the node receiving the getblocks request does not actually send
full blocks in response. The node sends an "inv" message containing just
the hashes of the series of blocks that fit the request, which verifies
that the node does indeed have blocks to send that the remote node does
not have (but does not presume the remote node wants the full blocks yet).
When the local node receives the "inv" message, it will request the actual
blocks with a "getdata" message. See Below.
But first, here is more detail how the remote node sends the "inv" message
in response to the "getblocks" request sent by the local node. The remote
node calls pFrom->PushInventory which is a method on the CNode instance that
represents the node that requested the blocks (the local node in this
walkthrough), and PushInventory adds the block hash to the vInventoryToSend
variable of the CNode. The SendMessages function in main.cpp will take the
inv items out of vInventoryToSend and add it to a vInv variable which
means they are really ready for sending.[4]
- side node -
The reason for the seperate variable is that some inventory items
(transactions only right now) may be "trickled" to the remote node,
which means they may kept from from being sent right away.
When the vInv variable fills up with 1000 entries, a message is queued
with those 1000 entries and the loop continues. At the end, any
remaining entries are sent in a final "inv" message.
When the local node receives the "inv" message, it will request the actual
block with a "getdata" message. To be precise, the node calls pfrom->AskFor
to request the block, and that method queues the request for the blocks in
mapAskFor, and the multipurpose SendMessage() sends the "getdata" requests
in batches of 1000 entries from the map.[5]
The code attempts to limit redundant requests to every 2 minutes for the
same block by using a map called mapAlreadyAskedFor to delay the message
if necessary.[6]
-- Block Batching --
The responding node to a "getblocks" request attempts to limit the response
to the requestor to 500 blocks.[7]
However, in a peculiar twist, if the requestor appears to have diverged
from the main branch, the node will send as many blocks as necessary to
replace the entire bad chain of the requestor, from the lastest common block
between the nodes, up to the last block the requestor has (on their bad main
branch). That is in addition to the 500 "catch up" blocks for main branch
updates that will also be sent.[8]
Commentary:
This begs the question: what is the purpose of the 500 limit batch code?
It cannot be to DoS protect the responder from sending too much, because
the requestor can easily claim they are diverged from the main branch and
the responder will attempt to send the entire chain. It cannot be to
protect the client because if the poor client had been fooled into
receiving a long fake chain, then why further punish them by DoSsing
them with the entire chain?
I think a possible assumption was made in thinking that a node off the
main branch will never be able to catch up if we limit the number of
blocks we send them to a number below the number of blocks they are
behind.[C1] However, that should not be the case. Eventually the
requestor should receive the main chain blocks as the "continue mechanism"
(see below) proceeds on to further blocks.
C1. https://github.com/bitcoin/bitcoin/commit/52f4cb48590a706caf7a492e8d94b85620d5cd33
Note that in addition to a flat limit on the number of blocks queued for
sending, bitcoind also limits the total size of the blocks that are being
queued. This is currently limited to half the send buffer size[9], which is
10MB, for a limit of 5MB of queued blocks for send.[10]
Commentary:
The 5MB send size limit is enough for about 200 blocks these
days (Aug2011). Batching based on send limits is what matters for
recent blocks since around block 127000 when the block size reached
10K and 500 blocks exceeds the size limit (10K*500=5MB).
The dominance of the size limit for all but the early blocks raises
the question of whether the 500 block limit has enough relevance to
remain in the code.
It was relevant for a period of time. The size limit [C1] came along long
after the 500 block limit [C2] and I can imagine there was no pressing
need to remove Satoshi's earlier code. No need to go mucking around stuff
with unknown side effects, perhaps.
C1. https://github.com/bitcoin/bitcoin/commit/497317453422611a077f7f195eb193d3bb597a9c
C2. https://github.com/bitcoin/bitcoin/commit/c5c7911dab8732861affbe66849a100da62f7464
-- Batch Continue Mechanism --
When a node is finished sending a batch of block inventory, it records the
hash of the last block in the batch.[11] When the node receives a request
for that full block, it realizes the remove node is done with the current
batch and directly queues a special "inv" message (bypassing the normal
SendMessage mechanism) with one block hash entry containing the
latest block hash.[12] When the remote node receives that "inv" message,
it will see that it does not have that block and it it will ask for that
block as described above. However, this time when it receives the block
and processes it, it will notice that it does not have the previous block,
so it will record the latest block as an "orphan" block, and will request a
block update starting with the latest block it has up to the block before
the orphan [13], in order to fill in the gap. That goes out as a "getblocks"
request and the whole batch process repeats itself.
However, there is a twist. When the next batch finishes, the remote node
sending the blocks will send the "inv" with latest block hash as usual, and
the local node will notice it already has this block in the orphan block
map this time and so it will skip requesting the block and directly ask
for a block update.[14] This process will continue until the last
block prior to the latest block is received. At the end of processing
that block, it will notice there is an orphan that pointed to this block
and will process the orphan block, (and any other orphans, recursively)
thus completing the entire process.[15]
-- Stall Recovery --
If the batching processes is interrupted for some reason, such as the
remote node failing to honor the "Batch Continue Mechanism" or if a
disconnection occurs, there is a way for the process to restart. When
a new block is solved and advertised around[16], any nodes that are
behind will notice the new block in the "inv" and that will trigger it
to request a "getblocks" update from the node that sent it the message.
That will cause blocks to be sent starting from wherever in the block
chain that the node that is behind is currently at.
-- Long Orphan Chains --
In various tests, it has proven relatively common (say more than one
in ten) to discover nodes that are significantly behind on the block
chain, probably because they are in the process of catching up as well.
Since a well connected node will have at least 8 and up to dozens of
connections, it is fairly likely that a new node will connect to
another node that is also catching up.
Nodes that are catching up will advertise the blocks they are processing,
as they accept blocks into their main chain, to every other node.[16]
While there is code to prevent advertising old blocks before a certain
checkpoint, that code also has a clause that does advertise blocks to
remote nodes if the block height is over the remote node's current best
height minus 2000 blocks.[17] This appears to allow nodes to "help" other
nodes catch up, even if they are both processing old blocks.
These advertisements cause the local node to request those blocks
from the remote node, which could be blocks well into the future compared
to what has been processed locally. Due to the way blocks are requested,
the remote node will send a large batch of blocks in response and will
continue sending blocks to the local node until it reaches the end.
Note that this is likely to occur at the same time the local node is
downloading earlier blocks on the main chain from another node. That
process may eventually catch up with the orphan chain and produce a
very, very long operation to revalidate and connect up all the orphan
blocks. Orphan chains over ten thousand blocks long, taking over an hour
to process are possible.
Therefore, two nodes talking to each other that are both catching up can
lead to suboptimal interactions, especially when one both are far behind
and one is far ahead of the other.
-- Flood Limit Effects --
Even with the batching mechanism described above, there are scenarios
that occur that result in the remote node overflowing the local receive
buffer while blocks are being exchanged.
For example, if a remote node is "catching up", it will advertise each block
it processes to the local node in certain circumstances (see above [17]).
The local node will request each of those blocks right away. There is no
protection against the local node requesting too many of these blocks.
The remote node will send all blocks requested. There is no protection
against the remote node sending too many blocks before the local node has
time to process them, in this circumstance.
The local receive buffer can fill up. When the local node notices a receive
buffer is full, it disconnects that node connection.[18]
If sets the fDisconnect flag, and once the buffers are empty[19], the
socket is closed.
-- Performance --
As of September 1, 2011, on a server class computer circa 2005 running
Ubuntu with a Comcast cable internet connection takes over 10 hours
to download and process the block chain. While it is debatable what
the bottleneck is early in the download process, it is clear from
the processing of recent blocks that the network is not the bottleneck
for all but the slowest internet connections.
Blocks are taking over a second, on average, to process once downloaded.[20]
However, the average size of a block is only around 24 kilobytes
in August 2011. It certainly does not take 1 second to download 24K.
Also, testing reveals very large queues of blocks being processed per
message loop, which is not what you would expect if the thread was
pulling them out of the queue as they arrive on the sockets.
There are a number of "false signals" that lead one to believe the problem
is with network performance. The first false signal is that, as of
August 2011, nearly all of the first 60 or 70% of blocks downloaded are
very small. Recent average block sizes are around one hundred times bigger!
So, almost all of a sudden, the block rate goes from very fast to very slow.
It looks like something went wrong. In reality, if you measure the rate
of block processing by kilobyte, the rate remains relatively constant.
Another false signal is related to the fact that message queues are
processed to completion, one at a time per node. This can result in big
backups of messages from other nodes. So, a long period of increasing
blocks may freeze for long periods as other nodes are serviced. Consider
that block downloads typically come from just one remote node (at
least until a miner or other relaying or downloading node advertises
a late block and disrupts the process) and so all the work might
be on one node. Things go fast processing the blocks from a node,
and then that looks like it stops as "addr" messages are processed from
other nodes and other work is done. But it looks like something is wrong.
Also, the orphaning effects described above can lead to excessive block
processing with nothing to show for it until the orphan chain is connected.
Also, you do ocassionally run into a node that is slow to respond, perhaps
because they are also processing blocks or because they have a slow machine
or connection.
All of the above contributes to heavy "jitter" in the block download process,
and that is a more frustrating user experience than a constant download rate.
Commentary:
The download delay has become much worse recently. Block sizes have
reached the point where it takes (using the once second avg) many
hours to process the chain and that time is increasing about an hour
every three weeks. So, if your laptop was off for three weeks, it
could take an hour to catch up. Initial download times are currently
set to increase 15 hours per year at todays transaction volume.
Clearly, downloading a preprocessed database of the block chain will
become an increasingly desirable option for many users.
Moreover, the current mechanism does not appear to be appropriate
for the casual user who runs the software part time on a laptop or home
computer, for example, and wants to have on-demand access to block
chain verification using the bitcoin protocol. Several solutions
proposed so far requires these users to place trust in something more
than the bitcoin network and protocol itself. Several "trust events" in
the bitcoin community in 2011 have probably hindered these solutions,
but in my opinion they are necessary.
Footnotes
1. See pfrom->fInbound where pfrom is a CNode.
2. See ProcessMessage() in main.cpp where strCommand == "version".
3. See CBlockLocator in main.h.
4. See Message: inventory in SendMessage in main.cpp.
5. See Message: getdata at the end of SendMessage in main.cpp.
6. See CNode::AskFor in net.h.
7. See ProcessMessage() in main.cpp where strCommand =="getblocks".
8. See int nLimit = 500 + locator.GetDistanceBack();
in ProcessMessage in main.cpp where strCommand =="getblocks".
9. See if (--nLimit <= 0 || nBytes >= SendBufferSize()/2)
in ProcessMessage() in main.cpp where strCommand =="getblocks".
10. See inline unsigned int SendBufferSize() {
return 1000*GetArg("-maxsendbuffer", 10*1000); }
in net.h.
11. See pfrom->hashContinue = pindex->GetBlockHash();
in ProcessMessage() in main.cpp where strCommand =="getblocks".
12. See: if (inv.hash == pfrom->hashContinue)
in ProcessMessage() in main.cpp where strCommand =="getdata".
13. See:
// Ask this guy to fill in what we're missing
if (pfrom)
pfrom->PushGetBlocks(pindexBest, GetOrphanRoot(pblock2));
in ProcessBlock() in main.cpp.
14. See:
else if (inv.type == MSG_BLOCK && mapOrphanBlocks.count(inv.hash))
pfrom->PushGetBlocks(pindexBest, GetOrphanRoot(mapOrphanBlocks[inv.hash]));
in ProcessMessage() in main.cpp where strCommand =="inv".
15. See:
// Recursively process any orphan blocks that depended on this one
in ProcessBlock() in main.cpp.
16. See the last block of code in AcceptBlock in main.cpp.
17. See:
if (nBestHeight > (pnode->nStartingHeight != -1 ? pnode->nStartingHeight - 2000 : 134444))
in AcceptBlock() in main.cpp.
18. See: if (nPos > ReceiveBufferSize()) {
in ThreadSocketHandler2() in net.cpp.
19. See:
if (pnode->fDisconnect ||
(pnode->GetRefCount() <= 0 && pnode->vRecv.empty() && pnode->vSend.empty()))
in ThreadSocketHandler2() in net.cpp.
20. This is from the authors experience and also
see https://bitcointalksearch.org/topic/any-way-to-get-client-startup-faster-31376.
--
Search on "Satoshi Client Operation" for more articles in this series.
Topic: Satoshi Client Operation: Block Exchange (Read 4195 times)
Jump to: