Performance Tuning
Overview
The Firedancer validator is composed of a handful of threads, each performing one of eleven distinct jobs. Some jobs only need one thread to do them, but certain jobs require many threads performing the same work in parallel.
Each thread is given a CPU core to run on, and threads take ownership of the core: never sleeping or letting the operating system use it for another purpose. The combination of a job, and the thread it runs on, and the CPU core it is assigned to is called a tile. The eleven kinds of tile are,
| Tile | Description |
|---|---|
net | Sends and receives network packets from the network device |
quic | Receives transactions from clients, performing all connection management and packet processing to manage and implement the QUIC protocol |
verify | Verifies the cryptographic signature of incoming transactions, filtering invalid ones |
dedup | Checks for and filters out duplicated incoming transactions |
pack | Collects incoming transactions and smartly schedules them for execution when we are leader |
bank | Executes transactions that have been scheduled when we are leader |
poh | Continuously hashes in the background, and mixes the hash in with executed transactions to prove passage of time |
shred | Distributes block data to the network when leader, and receives and retransmits block data when not leader |
store | Receives block data when we are leader, or from other nodes when they are leader, and stores it locally in a database on disk |
metric | Collects monitoring information about other tiles and serves it on a HTTP endpoint |
sign | Holds the validator private key, and receives and responds to signing requests from other tiles |
These tiles communicate with each other via shared memory queues. The work each tile performs and how they communicate with each other is fixed, but the count of each tile kind and which CPU cores they are assigned to is set by your configuration, and this the primary way to tune the performance of Firedancer.
Configuration
The default configuration provided if no options are specified is given in the default.toml file:
[layout]
affinity = "1-16"
agave_affinity = "17-31"
net_tile_count = 1
quic_tile_count = 1
verify_tile_count = 4
bank_tile_count = 2
shred_tile_count = 1Note that not all tiles have a configurable count. The dedup, pack, poh, store, metric, and sign tiles are fixed at one thread each.
The assignment of tiles to CPU cores is determined by the affinity string, which is documented fully in the default.toml file itself. The Frankendancer validator currently starts an Agave process to perform functionality like replay, gossip, and repair that is not yet implemented in Firedancer. The agave_affinity string determines the CPU cores that are given to the threads of this Agave process.
The following table shows the performance of the adjustable tiles on an Intel Icelake core, along with some performance notes and recommendations for mainnet-beta,
| Tile | Default | Notes |
|---|---|---|
net | 1 | Handles >1M TPS per tile. Designed to scale out for future network conditions, but there is no need to run more than 1 net tile at the moment on mainnet-beta |
quic | 1 | Handles >1M TPS per tile. Designed to scale out for future network conditions, but there is no need to run more than 1 QUIC tile at the moment on mainnet-beta |
verify | 4 | Handles 20-40k TPS per tile. Recommend running many verify tiles, as signature verification is the primary bottleneck of the application |
bank | 2 | Handles 20-40k TPS per tile, with diminishing returns from adding more tiles. Designed to scale out for future network conditions, but 2 tiles is enough to handle current mainnet-beta conditions. Can be increased further when benchmarking to test future network performance |
shred | 1 | Throughput is mainly dependent on cluster size, 1 tile is enough to handle current mainnet-beta conditions. In benchmarking, if the cluster size is small, 1 tile can handle >1M TPS |
Testing
Firedancer includes a simple benchmarking tool for measuring the transaction throughput of the validator when it is leader, in transactions per second (TPS). In practice, the Solana network performance is limited by two factors that are unrelated to what this tool measures:
- The replay performance of the slowest nodes in the network, and if they can keep up
- The consensus limits on block size and data size
In particular, consensus limits on the Solana protocol limit the network strictly to around 81,000 TPS. But the tool can be useful for testing local affinity and layout configurations.
The benchmark runs on a single machine and performs the following:
- A new genesis is created, and set of accounts are pre-funded
- A set of CPU cores is assigned to generating and signing simple transactions using these accounts as fast as possible
- Another set of CPU cores is assigned to sending these transfers via. QUIC over loopback to the locally running validator
- Around once a second, an RPC call is made to get the total count of transactions that have executed on the chain, and this information is printed to the console
The benchmark is currently quite synthetic, as it only measures single node performance, in an idealized case where all transactions are non-conflicting.
Running
The benchmark command is part of the fddev development binary, which can be built with make -j fddev. With the binary in hand, we can run our benchmark, here it will be on a 32 physical core AMD EPYC 7513:
$ lscpu
Architecture: x86_64
CPU(s): 64
On-line CPU(s) list: 0-63
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 1
NUMA node(s): 1
Vendor ID: AuthenticAMD
Model name: AMD EPYC 7513 32-Core Processor$ ./build/native/gcc/bin/fddev bench
NOTICE main configure.c(106): kill ... configuring
NOTICE main configure.c(81): netns ... skipping .. not enabled
NOTICE main configure.c(102): hugetlbfs ... already valid
NOTICE main configure.c(102): sysctl ... already valid
[ ... snip ... ]
NOTICE bencho:0 fd_bencho.c(137): 63373 txn/s
NOTICE bencho:0 fd_bencho.c(137): 63340 txn/sWe have not provided a configuration file to the bench command, so it is using the stock configuration from default.toml and reaching around 63,000 TPS.
Let's take a look at the performance with the monitor command and see if we can figure out what's going on.
$ ./build/native/gcc/bin/fddev monitor
snapshot for 2024-07-29 16:44:59.066296419 GMT+00
tile | pid | stale | heart | sig | in backp | backp cnt | % hkeep | % backp | % wait | % ovrnp | % ovrnr | % filt1 | % filt2 | % finish
---------+---------+------------+-------+------------+----------+---------------------+----------+----------+----------+----------+----------+----------+----------+----------
net | 2555236 | - | - | run( run) | -( -) | 0( +0) | 35.511 | 0.000 | 54.233 | 0.000 | 0.000 | 0.000 | 10.255 | 0.000
quic | 2555236 | - | - | run( run) | -( -) | 0( +0) | 0.405 | 0.000 | 95.501 | 0.000 | 0.000 | 0.000 | 0.000 | 4.094
verify | 2555236 | - | - | run( run) | -( -) | 0( +0) | 0.081 | 0.000 | 0.047 | 0.000 | 0.000 | 0.084 | 0.000 | 99.788
verify | 2555236 | - | - | run( run) | -( -) | 0( +0) | 0.082 | 0.000 | 0.036 | 0.000 | 0.000 | 0.078 | 0.000 | 99.804
verify | 2555236 | - | - | run( run) | -( -) | 0( +0) | 0.123 | 0.000 | 0.000 | 0.000 | 0.000 | 0.111 | 0.000 | 99.765
[ ... snip ... ]If we narrow in on just the verify tiles we can see the problem: all of the verify tiles are completely busy processing incoming transactions, and so additional transactions are being dropped. Here % finish indicates the percentage of time the tile is occupied doing work, while overnp cnt indicates that the tile is being overrun by the quic tile and dropping transactions.
tile | sig | in backp | backp cnt | % backp | % finish
---------+------------+----------+---------------+----------+----------
verify | run( run) | -( -) | 0( +0) | 0.000 | 99.788
verify | run( run) | -( -) | 0( +0) | 0.000 | 99.804
verify | run( run) | -( -) | 0( +0) | 0.000 | 99.765
verify | run( run) | -( -) | 0( +0) | 0.000 | 99.745
link | tot TPS | ovrnp cnt | ovrnr cnt | slow cnt | tx seq
------------------+----------+---------------------+---------------------+---------------------+-------------------
quic->verify | 17.2K | 9( +1) | 0( +0) | 0( +0) | 507134( +7149)
quic->verify | 17.2K | 9( +1) | 0( +0) | 0( +0) | 507134( +7149)
quic->verify | 11.8K | 15( +0) | 0( +0) | 0( +0) | 507134( +7149)
quic->verify | 17.0K | 9( +1) | 0( +0) | 0( +0) | 507134( +7149)This configuration is not ideal. With some tuning to increase the number of verify tiles, and a few other changes we can try to achieve a higher TPS rate,
[ledger]
# Place the ledger in memory rather than on disk so that writing the
# ledger is not a performance bottleneck
path = "/data/shm/{name}/ledger"
[layout]
# We will need a lot of verify tiles, and a few more bank tiles to be
# able to execute at higher TPS rates. Increase their core counts, and
# assign the tiles to cores. We only need 1 shred tile, since there is
# only 1 node in the cluster it can handle a high TPS rate by itself
affinity = "14-57,f1"
agave_affinity = "58-63"
verify_tile_count = 30
bank_tile_count = 6
shred_tile_count = 1
[development.genesis]
# The default amount of accounts to use for the benchmark is 1024, but
# to reach higher transaction throughput we need more accounts so that
# more transfers can be handled in parallel
fund_initial_accounts = 32768
[development.bench]
# benchg tiles are used to generate and sign transactions in the
# benchmarking tool, we are going to need more of them to test higher
# TPS rate
benchg_tile_count = 12
# benchs tiles are for sending the transactions to Firedancer over
# loopback, and we will need an extra one of these as well
benchs_tile_count = 2
# Assign these benchg, benchs (and the bencho tile which orchestrates
# the benchmarking) to some CPU cores. The bencho assignment is
# floating as it is not performance sensitive
affinity = "f1,0-13"
# The Solana protocol consensus limits restrict the benchmark to
# around 81,000 TPS. We have special options to increase these limits
# for testing and benchmarking
larger_max_cost_per_block = true
larger_shred_limits_per_block = true
[rpc]
# Tracking certain transaction history and metadata to serve RPC
# requests is expensive and can slow down our validator, turn this
# functionality off
transaction_history = false
extended_tx_metadata_storage = falseNow try running again,
$ ./build/native/gcc/bin/fddev bench --config src/app/fdctl/config/bench-zen3-32core.toml
[ ... snip ... ]
NOTICE bencho:0 fd_bencho.c(137): 191180 txn/s
NOTICE bencho:0 fd_bencho.c(137): 308027 txn/s
NOTICE bencho:0 fd_bencho.c(137): 12389 txn/s
NOTICE bencho:0 fd_bencho.c(137): 89564 txn/s
NOTICE bencho:0 fd_bencho.c(137): 92380 txn/sWe start out with a higher TPS rate but it quickly falls back to around 90k TPS. We can try to figure out why by running the monitor command.
$ ./build/native/gcc/bin/fddev monitor --config src/app/fdctl/config/bench-zen3-32core.toml
snapshot for 2024-07-29 16:44:59.066296419 GMT+00
tile | pid | stale | heart | sig | in backp | backp cnt | % hkeep | % backp | % wait | % ovrnp | % ovrnr | % filt1 | % filt2 | % finish
---------+---------+------------+-------+------------+----------+---------------------+----------+----------+----------+----------+----------+----------+----------+----------
net | 2239972 | - | - | run( run) | -( -) | 0( +0) | 48.576 | 0.000 | 51.424 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000
quic | 2239972 | - | - | run( run) | -( -) | 0( +0) | 0.553 | 0.000 | 88.238 | 0.000 | 0.000 | 0.000 | 0.000 | 11.208
verify | 2239972 | - | - | run( run) | -( -) | 0( +0) | 0.199 | 0.000 | 0.000 | 0.000 | 0.000 | 0.804 | 0.000 | 98.997
verify | 2239972 | - | - | run( run) | -( -) | 0( +0) | 0.177 | 0.000 | 0.000 | 0.000 | 0.000 | 0.796 | 0.000 | 99.027
verify | 2239972 | - | - | run( run) | -( -) | 0( +0) | 0.177 | 0.000 | 0.000 | 0.000 | 0.000 | 0.796 | 0.000 | 99.027
[ ... snip ... ]The culprit is visible in the output, which will be clearer if we filter it down to the relevant information,
tile | sig | in backp | backp cnt | % backp | % finish
---------+------------+----------+---------------+----------+----------
poh | run( run) | err(err) | 1002( +22) | 76.918 | 7.892
shred | run( run) | err(err) | 1462( +22) | 95.857 | 3.386
store | run( run) | -( -) | 0( +0) | 0.000 | 99.973Here we see what is happening. The blockstore is completely busy spending 99.973% of its time storing data, while the PoH and shred tiles are in back-pressure waiting for the blockstore to catch up. The blockstore is an Agave component built on RocksDB that is not rewritten as part of Frankendancer.
[development.bench]
disable_blockstore_from_slot = 1We can disable the blockstore specifically for benchmarking, to show the performance of just the Firedancer components in the leader pipeline. Now we can run one more time and see a reasonably good value for the TPS throughput of Firedancer on this machine.
$ ./build/native/gcc/bin/fddev bench --config src/app/fdctl/config/bench-zen3-32core.toml
[ ... snip ... ]
NOTICE bencho:0 fd_bencho.c(137): 272840 txn/s
NOTICE bencho:0 fd_bencho.c(137): 278380 txn/s
NOTICE bencho:0 fd_bencho.c(137): 268246 txn/s