add metrics for rdb lru cache (#2586)
This is a first step towards measuring the efficiency of the LRU caches
over time - metrics can be collected during import or when running
regulary.
Since `nim-metrics` carries some overhead for its default way of
reporting metrics, this PR implements a custom collector over atomic
counters, given that this is one of the hottest spots in the block
processing pipeline.
Using a compile-time flag, the same metrics can be printed on exit which
is useful when comparing different strategies for caching - here's a
recent run over blocks 16000001-1616384 - this is a good candidate to
expose in a better way in the future, maybe:
```
state vtype miss hit total hitrate
Account Leaf 4909417 4466215 9375632 47.64%
Account Branch 20742574 72015123 92757697 77.64%
World Leaf 940483 1140946 2081429 54.82%
World Branch 8224151 131496580 139720731 94.11%
all all 34816625 209118864 243935489 85.73%
```
This commit is contained in:
Jacek Sieka
GitHub
parent
ef1bab0802
commit
35cc78c86d
|
|
@ -20,7 +20,9 @@ import
|
|||
stew/keyed_queue,
|
||||
../../[aristo_blobify, aristo_desc],
|
||||
../init_common,
|
||||
./rdb_desc
|
||||
./rdb_desc,
|
||||
metrics,
|
||||
std/concurrency/atomics
|
||||
|
||||
const
|
||||
extraTraceMessages = false
|
||||
|
|
@ -33,6 +35,73 @@ when extraTraceMessages:
|
|||
logScope:
|
||||
topics = "aristo-rocksdb"
|
||||
|
||||
type
|
||||
RdbVtxLruCounter = ref object of Counter
|
||||
RdbKeyLruCounter = ref object of Counter
|
||||
|
||||
LruCounter = array[bool, Atomic[uint64]]
|
||||
|
||||
StateType = enum
|
||||
Account
|
||||
World
|
||||
|
||||
var
|
||||
# Hit/miss counters for LRU cache - global so as to integrate easily with
|
||||
# nim-metrics and `uint64` to ensure that increasing them is fast - collection
|
||||
# happens from a separate thread.
|
||||
# TODO maybe turn this into more general framework for LRU reporting since
|
||||
# we have lots of caches of this sort
|
||||
rdbVtxLruStats: array[StateType, array[VertexType, LruCounter]]
|
||||
rdbKeyLruStats: array[StateType, LruCounter]
|
||||
|
||||
var
|
||||
rdbVtxLruStatsMetric {.used.} = RdbVtxLruCounter.newCollector(
|
||||
"aristo_rdb_vtx_lru_total",
|
||||
"Vertex LRU lookup (hit/miss, world/account, branch/leaf)",
|
||||
labels = ["state", "vtype", "hit"],
|
||||
)
|
||||
rdbKeyLruStatsMetric {.used.} = RdbKeyLruCounter.newCollector(
|
||||
"aristo_rdb_key_lru_total", "HashKey LRU lookup", labels = ["state", "hit"]
|
||||
)
|
||||
|
||||
template to(v: RootedVertexID, T: type StateType): StateType =
|
||||
if v.root == VertexID(1): StateType.World else: StateType.Account
|
||||
|
||||
template inc(v: var LruCounter, hit: bool) =
|
||||
discard v[hit].fetchAdd(1, moRelaxed)
|
||||
|
||||
template get(v: LruCounter, hit: bool): uint64 =
|
||||
v[hit].load(moRelaxed)
|
||||
|
||||
method collect*(collector: RdbVtxLruCounter, output: MetricHandler) =
|
||||
let timestamp = collector.now()
|
||||
|
||||
# We don't care about synchronization between each type of metric or between
|
||||
# the metrics thread and others since small differences like this don't matter
|
||||
for state in StateType:
|
||||
for vtype in VertexType:
|
||||
for hit in [false, true]:
|
||||
output(
|
||||
name = "aristo_rdb_vtx_lru_total",
|
||||
value = float64(rdbVtxLruStats[state][vtype].get(hit)),
|
||||
labels = ["state", "vtype", "hit"],
|
||||
labelValues = [$state, $vtype, $ord(hit)],
|
||||
timestamp = timestamp,
|
||||
)
|
||||
|
||||
method collect*(collector: RdbKeyLruCounter, output: MetricHandler) =
|
||||
let timestamp = collector.now()
|
||||
|
||||
for state in StateType:
|
||||
for hit in [false, true]:
|
||||
output(
|
||||
name = "aristo_rdb_key_lru_total",
|
||||
value = float64(rdbKeyLruStats[state].get(hit)),
|
||||
labels = ["state", "hit"],
|
||||
labelValues = [$state, $ord(hit)],
|
||||
timestamp = timestamp,
|
||||
)
|
||||
|
||||
# ------------------------------------------------------------------------------
|
||||
# Public functions
|
||||
# ------------------------------------------------------------------------------
|
||||
|
|
@ -60,8 +129,11 @@ proc getKey*(
|
|||
# Try LRU cache first
|
||||
var rc = rdb.rdKeyLru.lruFetch(rvid.vid)
|
||||
if rc.isOK:
|
||||
rdbKeyLruStats[rvid.to(StateType)].inc(true)
|
||||
return ok(move(rc.value))
|
||||
|
||||
rdbKeyLruStats[rvid.to(StateType)].inc(false)
|
||||
|
||||
# Otherwise fetch from backend database
|
||||
# A threadvar is used to avoid allocating an environment for onData
|
||||
var res{.threadvar.}: Opt[HashKey]
|
||||
|
|
@ -90,6 +162,7 @@ proc getVtx*(
|
|||
# Try LRU cache first
|
||||
var rc = rdb.rdVtxLru.lruFetch(rvid.vid)
|
||||
if rc.isOK:
|
||||
rdbVtxLruStats[rvid.to(StateType)][rc.value().vType].inc(true)
|
||||
return ok(move(rc.value))
|
||||
|
||||
# Otherwise fetch from backend database
|
||||
|
|
@ -105,13 +178,62 @@ proc getVtx*(
|
|||
return err((errSym,error))
|
||||
|
||||
if not gotData:
|
||||
res.ok(VertexRef(nil))
|
||||
elif res.isErr():
|
||||
# As a hack, we count missing data as leaf nodes
|
||||
rdbVtxLruStats[rvid.to(StateType)][VertexType.Leaf].inc(false)
|
||||
return ok(VertexRef(nil))
|
||||
|
||||
if res.isErr():
|
||||
return err((res.error(), "Parsing failed")) # Parsing failed
|
||||
|
||||
rdbVtxLruStats[rvid.to(StateType)][res.value().vType].inc(false)
|
||||
|
||||
# Update cache and return
|
||||
ok rdb.rdVtxLru.lruAppend(rvid.vid, res.value(), RdVtxLruMaxSize)
|
||||
|
||||
# ------------------------------------------------------------------------------
|
||||
# End
|
||||
# ------------------------------------------------------------------------------
|
||||
|
||||
when defined(printStatsAtExit):
|
||||
# Useful hack for printing exact metrics to compare runs with different
|
||||
# settings
|
||||
import std/[exitprocs, strformat]
|
||||
addExitProc(
|
||||
proc() =
|
||||
block vtx:
|
||||
var misses, hits: uint64
|
||||
echo "vtxLru(", RdVtxLruMaxSize, ")"
|
||||
echo " state vtype miss hit total hitrate"
|
||||
for state in StateType:
|
||||
for vtype in VertexType:
|
||||
let
|
||||
(miss, hit) = (
|
||||
rdbVtxLruStats[state][vtype].get(false),
|
||||
rdbVtxLruStats[state][vtype].get(true),
|
||||
)
|
||||
hitRate = float64(hit * 100) / (float64(hit + miss))
|
||||
misses += miss
|
||||
hits += hit
|
||||
echo &"{state:>8} {vtype:>8} {miss:>10} {hit:>10} {miss+hit:>10} {hitRate:>6.2f}%"
|
||||
let hitRate = float64(hits * 100) / (float64(hits + misses))
|
||||
echo &" all all {misses:>10} {hits:>10} {misses+hits:>10} {hitRate:>6.2f}%"
|
||||
|
||||
block key:
|
||||
var misses, hits: uint64
|
||||
echo "keyLru(", RdKeyLruMaxSize, ") "
|
||||
|
||||
echo " state miss hit total hitrate"
|
||||
|
||||
for state in StateType:
|
||||
let
|
||||
(miss, hit) =
|
||||
(rdbKeyLruStats[state].get(false), rdbKeyLruStats[state].get(true))
|
||||
hitRate = float64(hit * 100) / (float64(hit + miss))
|
||||
misses += miss
|
||||
hits += hit
|
||||
|
||||
echo &"{state:>8} {miss:>10} {hit:>10} {miss+hit:>10} {hitRate:>5.2f}%"
|
||||
|
||||
let hitRate = float64(hits * 100) / (float64(hits + misses))
|
||||
echo &" all {misses:>10} {hits:>10} {misses+hits:>10} {hitRate:>5.2f}%"
|
||||
)
|
||||
|
|
|
|||
Loading…
Reference in New Issue