nimbus-eth1/nimbus/sync/snap/worker/ticker.nim

200 lines
6.1 KiB
Nim
Raw Normal View History

# Nimbus - Fetch account and storage states from peers efficiently
#
# Copyright (c) 2021 Status Research & Development GmbH
# Licensed under either of
# * Apache License, version 2.0, ([LICENSE-APACHE](LICENSE-APACHE) or
# http://www.apache.org/licenses/LICENSE-2.0)
# * MIT license ([LICENSE-MIT](LICENSE-MIT) or
# http://opensource.org/licenses/MIT)
# at your option. This file may not be copied, modified, or distributed
# except according to those terms.
import
std/[strformat, strutils],
chronos,
chronicles,
eth/[common, p2p],
stint,
../../../utils/prettify,
../../misc/timer_helper
{.push raises: [Defect].}
logScope:
topics = "snap-tick"
type
TickerStats* = object
pivotBlock*: Option[BlockNumber]
nAccounts*: (float,float) ## mean and standard deviation
accountsFill*: (float,float,float) ## mean, standard deviation, merged total
nAccountStats*: (int,int) ## #chunks, #dangling/missing nodes
nSlotLists*: (float,float) ## mean and standard deviation
nStorageQueue*: Option[int]
Prep for full sync after snap (#1253) * Split fetch accounts into sub-modules details: There will be separated modules for accounts snapshot, storage snapshot, and healing for either. * Allow to rebase pivot before negotiated header why: Peers seem to have not too many snapshots available. By setting back the pivot block header slightly, the chances might be higher to find more peers to serve this pivot. Experiment on mainnet showed that setting back too much (tested with 1024), the chances to find matching snapshot peers seem to decrease. * Add accounts healing * Update variable/field naming in `worker_desc` for readability * Handle leaf nodes in accounts healing why: There is no need to fetch accounts when they had been added by the healing process. On the flip side, these accounts must be checked for storage data and the batch queue updated, accordingly. * Reorganising accounts hash ranges batch queue why: The aim is to formally cover as many accounts as possible for different pivot state root environments. Formerly, this was tried by starting the accounts batch queue at a random value for each pivot (and wrapping around.) Now, each pivot environment starts with an interval set mutually disjunct from any interval set retrieved with other pivot state roots. also: Stop fishing for more pivots in `worker` if 100% download is reached * Reorganise/update accounts healing why: Error handling was wrong and the (math. complexity of) whole process could be better managed. details: Much of the algorithm is now documented at the top of the file `heal_accounts.nim`
2022-10-08 17:20:50 +00:00
nQueues*: int
TickerStatsUpdater* =
proc: TickerStats {.gcsafe, raises: [Defect].}
TickerRef* = ref object
## Account fetching state that is shared among all peers.
nBuddies: int
lastStats: TickerStats
statsCb: TickerStatsUpdater
logTicker: TimerCallback
started: Moment
visited: Moment
const
tickerStartDelay = chronos.milliseconds(100)
tickerLogInterval = chronos.seconds(1)
tickerLogSuppressMax = chronos.seconds(100)
# ------------------------------------------------------------------------------
# Private functions: pretty printing
# ------------------------------------------------------------------------------
Prep for full sync after snap make 6 (#1291) * Update log ticker, using time interval rather than ticker count why: Counting and logging ticker occurrences is inherently imprecise. So time intervals are used. * Use separate storage tables for snap sync data * Left boundary proof update why: Was not properly implemented, yet. * Capture pivot in peer worker (aka buddy) tasks why: The pivot environment is linked to the `buddy` descriptor. While there is a task switch, the pivot may change. So it is passed on as function argument `env` rather than retrieved from the buddy at the start of a sub-function. * Split queues `fetchStorage` into `fetchStorageFull` and `fetchStoragePart` * Remove obsolete account range returned from `GetAccountRange` message why: Handler returned the wrong right value of the range. This range was for convenience, only. * Prioritise storage slots if the queue becomes large why: Currently, accounts processing is prioritised up until all accounts are downloaded. The new prioritisation has two thresholds for + start processing storage slots with a new worker + stop account processing and switch to storage processing also: Provide api for `SnapTodoRanges` pair of range sets in `worker_desc.nim` * Generalise left boundary proof for accounts or storage slots. why: Detailed explanation how this works is documented with `snapdb_accounts.importAccounts()`. Instead of enforcing a left boundary proof (which is still the default), the importer functions return a list of `holes` (aka node paths) found in the argument ranges of leaf nodes. This in turn is used by the book keeping software for data download. * Forgot to pass on variable in function wrapper also: + Start healing not before 99% accounts covered (previously 95%) + Logging updated/prettified
2022-11-08 18:56:04 +00:00
# proc ppMs*(elapsed: times.Duration): string
# {.gcsafe, raises: [Defect, ValueError]} =
# result = $elapsed.inMilliseconds
# let ns = elapsed.inNanoseconds mod 1_000_000 # fraction of a milli second
# if ns != 0:
# # to rounded deca milli seconds
# let dm = (ns + 5_000i64) div 10_000i64
# result &= &".{dm:02}"
# result &= "ms"
#
# proc ppSecs*(elapsed: times.Duration): string
# {.gcsafe, raises: [Defect, ValueError]} =
# result = $elapsed.inSeconds
# let ns = elapsed.inNanoseconds mod 1_000_000_000 # fraction of a second
# if ns != 0:
# # round up
# let ds = (ns + 5_000_000i64) div 10_000_000i64
# result &= &".{ds:02}"
# result &= "s"
#
# proc ppMins*(elapsed: times.Duration): string
# {.gcsafe, raises: [Defect, ValueError]} =
# result = $elapsed.inMinutes
# let ns = elapsed.inNanoseconds mod 60_000_000_000 # fraction of a minute
# if ns != 0:
# # round up
# let dm = (ns + 500_000_000i64) div 1_000_000_000i64
# result &= &":{dm:02}"
# result &= "m"
#
# proc pp(d: times.Duration): string
# {.gcsafe, raises: [Defect, ValueError]} =
# if 40 < d.inSeconds:
# d.ppMins
# elif 200 < d.inMilliseconds:
# d.ppSecs
# else:
# d.ppMs
proc pc99(val: float): string =
if 0.99 <= val and val < 1.0: "99%"
elif 0.0 < val and val <= 0.01: "1%"
else: val.toPC(0)
# ------------------------------------------------------------------------------
# Private functions: ticking log messages
# ------------------------------------------------------------------------------
template noFmtError(info: static[string]; code: untyped) =
try:
code
except ValueError as e:
raiseAssert "Inconveivable (" & info & "): " & e.msg
proc setLogTicker(t: TickerRef; at: Moment) {.gcsafe.}
proc runLogTicker(t: TickerRef) {.gcsafe.} =
let
data = t.statsCb()
now = Moment.now()
if data != t.lastStats or tickerLogSuppressMax < (now - t.visited):
t.lastStats = data
t.visited = now
var
nAcc, nSto, bulk: string
pivot = "n/a"
Prep for full sync after snap make 4 (#1282) * Re-arrange fetching storage slots in batch module why; Previously, fetching partial slot ranges first has a chance of terminating the worker peer 9due to network error) while there were many inheritable storage slots on the queue. Now, inheritance is checked first, then full slot ranges and finally partial ranges. * Update logging * Bundled node information for healing into single object `NodeSpecs` why: Previously, partial paths and node keys were kept in separate variables. This approach was error prone due to copying/reassembling function argument objects. As all partial paths, keys, and node data types are more or less handled as `Blob`s over the network (using Eth/6x, or Snap/1) it makes sense to hold these `Blob`s as named field in a single object (even if not all fields are active for the current purpose.) * For good housekeeping, using `NodeKey` type only for account keys why: previously, a mixture of `NodeKey` and `Hash256` was used. Now, only state or storage root keys use the `Hash256` type. * Always accept latest pivot (and not a slightly older one) why; For testing it was tried to use a slightly older pivot state root than available. Some anecdotal tests seemed to suggest an advantage so that more peers are willing to serve on that older pivot. But this could not be confirmed in subsequent tests (still anecdotal, though.) As a side note, the distance of the latest pivot to its predecessor is at least 128 (or whatever the constant `minPivotBlockDistance` is assigned to.) * Reshuffle name components for some file and function names why: Clarifies purpose: "storages" becomes: "storage slots" "store" becomes: "range fetch" * Stash away currently unused modules in sub-folder named "notused"
2022-10-27 13:49:28 +00:00
nStoQue = "n/a"
let
Prep for full sync after snap make 6 (#1291) * Update log ticker, using time interval rather than ticker count why: Counting and logging ticker occurrences is inherently imprecise. So time intervals are used. * Use separate storage tables for snap sync data * Left boundary proof update why: Was not properly implemented, yet. * Capture pivot in peer worker (aka buddy) tasks why: The pivot environment is linked to the `buddy` descriptor. While there is a task switch, the pivot may change. So it is passed on as function argument `env` rather than retrieved from the buddy at the start of a sub-function. * Split queues `fetchStorage` into `fetchStorageFull` and `fetchStoragePart` * Remove obsolete account range returned from `GetAccountRange` message why: Handler returned the wrong right value of the range. This range was for convenience, only. * Prioritise storage slots if the queue becomes large why: Currently, accounts processing is prioritised up until all accounts are downloaded. The new prioritisation has two thresholds for + start processing storage slots with a new worker + stop account processing and switch to storage processing also: Provide api for `SnapTodoRanges` pair of range sets in `worker_desc.nim` * Generalise left boundary proof for accounts or storage slots. why: Detailed explanation how this works is documented with `snapdb_accounts.importAccounts()`. Instead of enforcing a left boundary proof (which is still the default), the importer functions return a list of `holes` (aka node paths) found in the argument ranges of leaf nodes. This in turn is used by the book keeping software for data download. * Forgot to pass on variable in function wrapper also: + Start healing not before 99% accounts covered (previously 95%) + Logging updated/prettified
2022-11-08 18:56:04 +00:00
accCov = data.accountsFill[0].pc99 &
"(" & data.accountsFill[1].pc99 & ")" &
"/" & data.accountsFill[2].pc99 &
"~" & data.nAccountStats[0].uint.toSI &
"/" & data.nAccountStats[1].uint.toSI
buddies = t.nBuddies
# With `int64`, there are more than 29*10^10 years range for seconds
up = (now - t.started).seconds.uint64.toSI
mem = getTotalMem().uint.toSI
noFmtError("runLogTicker"):
if data.pivotBlock.isSome:
Prep for full sync after snap (#1253) * Split fetch accounts into sub-modules details: There will be separated modules for accounts snapshot, storage snapshot, and healing for either. * Allow to rebase pivot before negotiated header why: Peers seem to have not too many snapshots available. By setting back the pivot block header slightly, the chances might be higher to find more peers to serve this pivot. Experiment on mainnet showed that setting back too much (tested with 1024), the chances to find matching snapshot peers seem to decrease. * Add accounts healing * Update variable/field naming in `worker_desc` for readability * Handle leaf nodes in accounts healing why: There is no need to fetch accounts when they had been added by the healing process. On the flip side, these accounts must be checked for storage data and the batch queue updated, accordingly. * Reorganising accounts hash ranges batch queue why: The aim is to formally cover as many accounts as possible for different pivot state root environments. Formerly, this was tried by starting the accounts batch queue at a random value for each pivot (and wrapping around.) Now, each pivot environment starts with an interval set mutually disjunct from any interval set retrieved with other pivot state roots. also: Stop fishing for more pivots in `worker` if 100% download is reached * Reorganise/update accounts healing why: Error handling was wrong and the (math. complexity of) whole process could be better managed. details: Much of the algorithm is now documented at the top of the file `heal_accounts.nim`
2022-10-08 17:20:50 +00:00
pivot = &"#{data.pivotBlock.get}/{data.nQueues}"
nAcc = (&"{(data.nAccounts[0]+0.5).int64}" &
&"({(data.nAccounts[1]+0.5).int64})")
nSto = (&"{(data.nSlotLists[0]+0.5).int64}" &
&"({(data.nSlotLists[1]+0.5).int64})")
Prep for full sync after snap make 4 (#1282) * Re-arrange fetching storage slots in batch module why; Previously, fetching partial slot ranges first has a chance of terminating the worker peer 9due to network error) while there were many inheritable storage slots on the queue. Now, inheritance is checked first, then full slot ranges and finally partial ranges. * Update logging * Bundled node information for healing into single object `NodeSpecs` why: Previously, partial paths and node keys were kept in separate variables. This approach was error prone due to copying/reassembling function argument objects. As all partial paths, keys, and node data types are more or less handled as `Blob`s over the network (using Eth/6x, or Snap/1) it makes sense to hold these `Blob`s as named field in a single object (even if not all fields are active for the current purpose.) * For good housekeeping, using `NodeKey` type only for account keys why: previously, a mixture of `NodeKey` and `Hash256` was used. Now, only state or storage root keys use the `Hash256` type. * Always accept latest pivot (and not a slightly older one) why; For testing it was tried to use a slightly older pivot state root than available. Some anecdotal tests seemed to suggest an advantage so that more peers are willing to serve on that older pivot. But this could not be confirmed in subsequent tests (still anecdotal, though.) As a side note, the distance of the latest pivot to its predecessor is at least 128 (or whatever the constant `minPivotBlockDistance` is assigned to.) * Reshuffle name components for some file and function names why: Clarifies purpose: "storages" becomes: "storage slots" "store" becomes: "range fetch" * Stash away currently unused modules in sub-folder named "notused"
2022-10-27 13:49:28 +00:00
if data.nStorageQueue.isSome:
nStoQue = $data.nStorageQueue.unsafeGet
info "Snap sync statistics",
up, buddies, pivot, nAcc, accCov, nSto, nStoQue, mem
t.setLogTicker(Moment.fromNow(tickerLogInterval))
proc setLogTicker(t: TickerRef; at: Moment) =
if not t.logTicker.isNil:
t.logTicker = safeSetTimer(at, runLogTicker, t)
# ------------------------------------------------------------------------------
# Public constructor and start/stop functions
# ------------------------------------------------------------------------------
proc init*(T: type TickerRef; cb: TickerStatsUpdater): T =
## Constructor
T(statsCb: cb)
proc start*(t: TickerRef) =
## Re/start ticker unconditionally
#debug "Started ticker"
t.logTicker = safeSetTimer(Moment.fromNow(tickerStartDelay), runLogTicker, t)
if t.started == Moment.default:
t.started = Moment.now()
proc stop*(t: TickerRef) =
## Stop ticker unconditionally
t.logTicker = nil
#debug "Stopped ticker"
# ------------------------------------------------------------------------------
# Public functions
# ------------------------------------------------------------------------------
proc startBuddy*(t: TickerRef) =
## Increment buddies counter and start ticker unless running.
if t.nBuddies <= 0:
t.nBuddies = 1
t.start()
else:
t.nBuddies.inc
proc stopBuddy*(t: TickerRef) =
## Decrement buddies counter and stop ticker if there are no more registered
## buddies.
t.nBuddies.dec
if t.nBuddies <= 0:
t.stop()
# ------------------------------------------------------------------------------
# End
# ------------------------------------------------------------------------------