805f9aa501
* tracing, write bench1() result into files * add context to computeTxEnv, remote_db cursor lazy cache allocation, * more results * added * added * added * more info * make linter happy * exit from tracing gorutines * remove unused param * re-run CI * re-run CI * remove concept txHandle * batch cursor in reports * remove debug output * fix shadow variables and unhandled errors
3.3 KiB
Run
go run ./cmd/rpctest --action=bench1 will print tmpDir.
And create in tmpDir results_*.csv and vegeta_*.txt files.
Command takes long time. Kill it when vegeta_turbo_geth_debug_storageRangeAt.txt is few MB.
File vegeta_turbo_geth_*.txt will produce load to turbo_geth node, vegeta_geth_*.txt to geth.
Change host/port in cmd/rpctest/main.go:routes variable.
By default go run ./cmd/rpctest --action=bench1 calling only Geth node
because cmd/rpctest/main.go:main calling it with first parameter needCompare=false.
Set needCompare=true to call Geth and TurboGeth nodes and compare results.
Install Vegeta
go get -u github.com/rs/jplot
go get -u github.com/rs/jaggr
go get -u github.com/tsenart/vegeta
Run vegeta
tmpDir = "/var/folders/x_/1mnbt25s3291zr5_fxhjfnq9n86kng/T/"
cat $(tmpDir)/turbo_geth_stress_test/vegeta_geth_debug_storageRangeAt.csv | vegeta attack -rate=600 -format=json -duration=20s | vegeta plot > plot.html
open plot.html
Mac environment changes
Results from my Macbook:
start rpcdaemon with turbo_geth:
GODEBUG=remotedb.debug=1 go run ./cmd/geth --remote-db-listen-addr localhost:9997 --rpcport 8545 --rpc --rpcapi eth,debug --nodiscover
GODEBUG=remotedb.debug=1 go run ./cmd/rpcdaemon --rpcapi eth,debug --rpcport 9545 --remote-db-addr 127.0.0.1:9997
On simple requests eth_getBlockByNumber RPC Daemon looks well:
cat /tmp/turbo_geth_stress_test/vegeta_turbo_geth_eth_getBlockByNumber.txt | vegeta attack -rate=1000 -format=json -duration=20s | vegeta report
300rps:
- Geth Alone: 80% of CPU, 95-Latency 2ms
- Geth Behind RPC Daemon: 25% of CPU
- RPC Daemon: 45% of CPU, 95-Latency 3ms
1000rps:
- Geth Alone: 200% of CPU, 95-Latency 3ms
- Geth Behind RPC Daemon: 50% of CPU
- RPC Daemon: 120% of CPU, 95-Latency 6ms
2000rps:
- Geth Alone: 400% of CPU, 95-Latency 15ms
- Geth Behind RPC Daemon: 100% of CPU
- RPC Daemon: 250% of CPU, 95-Latency 10ms
On complex request - debug_storageRangeAt producing >600 db.View calls with twice more .Bucket/.Cursor/.Seek calls:
echo "POST http://localhost:9545 \n Content-Type: application/json \n @$(pwd)/cmd/rpctest/heavyStorageRangeAt.json" | vegeta attack -rate=20 -duration=20s | vegeta report
10rps, batchSize 10K:
- Geth Alone: 100% of CPU, 95-Latency 15ms
- Geth Behind RPC Daemon: 200% of CPU
- RPC Daemon: 230% of CPU, 95-Latency 7s
10rps, batchSize 10:
- Geth Alone: 100% of CPU, 95-Latency 15ms
- Geth Behind RPC Daemon: 110% of CPU
- RPC Daemon: 100% of CPU, 95-Latency 230ms
Reason is: often usage of .GetAsOf() - this method does much .Next() and .Seek() calls.
Each .Seek() call invalidate internal batch cache of .Next() method and remote_db does read CursorBatchSize amount of keys again.
PoolSize=128, CursorBatchSize=10K -> 95-Latency 30s (eat all conns in pool)
PoolSize=128, CursorBatchSize=1K -> 95-Latency 6s (eat 50 conns in pool)
PoolSize=128, CursorBatchSize=100 -> 95-Latency 600ms (eat 5 conns in pool)
RPC daemon known problems:
- we call .Encode for each cmd/key/value - "maybe" it does syscall to send data. Maybe need buffered io.Writer or encode struct{cmd,key,value} at once. Need to write benchmark.