This PR is a more advanced form of the dirty-to-clean cacher (#18995),
where we reuse previous database write batches as datasets to uncache,
saving a dirty-trie-iteration and a dirty-trie-rlp-reencoding per block.
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.
receipts may be null for very short time in some condition. For this case, we should not add the null value into cache. Because you will not get the right result if you keep requesting that receipt.
Until this commit, when sending an RPC request that called `NewEVM`, a blank `vm.Config`
would be taken so as to set some options, based on the default configuration. If some extra
configuration switches were passed to the blockchain, those would be ignored.
This PR adds a function to get the config from the blockchain, and this is what is now used
for RPC calls.
Some subsequent changes need to be made, see https://github.com/ethereum/go-ethereum/pull/17955#pullrequestreview-182237244
for the details of the discussion.
The current trie memory database/cache that we do pruning on stores
trie nodes as binary rlp encoded blobs, and also stores the node
relationships/references for GC purposes. However, most of the trie
nodes (everything apart from a value node) is in essence just a
collection of references.
This PR switches out the RLP encoded trie blobs with the
collapsed-but-not-serialized trie nodes. This permits most of the
references to be recovered from within the node data structure,
avoiding the need to track them a second time (expensive memory wise).
Talk about "state" instead of "trie timing", "trie memory" and remove
the overzealous warning when the limit is just reached. Since the time
limit is always reached on slow machines, move the message to info level
so users don't freak out about internal details.
