Ethereum in numbers

Bane of Ethereum: Merkle Patricia trie

👋 Merkle tree containing account data leaves, linked together via 16-child internal nodes. 👋

Catch: The more accounts there are, the deeper the state trie becomes.

Bane of Ethereum: State trie depth

Logarithmic depth surely doesn't matter?

• Ethereum has 174M accounts ⇒ 6.85 internal depth + 1 leaf layer
• Plain transfers update 2 accounts ⇒ 15 new nodes in the account trie
• LevelDB stores data in 7 disk layers ⇒ amplifies at worse into 105 writes
• Old path read for root hash calculation ⇒ bumps to potentially 210 IO ops
• Mined blocks need to propagate ⇒ 210 ops miner side, 210 ops full node side

HDD                    capped at        80 IOPS  ⇒  0.19 TPS  (x2/3 = 0.12 TPS with disk pruning)
SSD (SATA 6)           capped at    90.000 IOPS  ⇒   214 TPS  (x2/3 =  142 TPS with disk pruning)
SSD (NVMe over PCIe 3) capped at   360.000 IOPS  ⇒   857 TPS  (x2/3 =  571 TPS with disk pruning)
SSD (NVMe over PCIe 4) capped at 1.000.000 IOPS  ⇒  2381 TPS  (x2/3 = 1587 TPS with disk pruning)

*Purely disk latency bounds, no in-memory optimizations (archive), no caching (small RAM)

Bane of Ethereum: State trie size

To raise the TPS, we must lower the disk IOPS:

• Keep things in memory and avoid hitting the disk
• OS uses free memory as disk cache ⇒ db shuffling in RAM
• Geth does in-memory pruning ⇒ ephemeral state never hits the db

Unfortunately, system memory is limited

• State outgrows the RAM ⇒ db writes revert to physical disk writes
• State becomes bigger ⇒ pruning needs more RAM or it flushes more

Bane of Ethereum: State trie growth

How fast is Ethereum's state growing (5th June, '22 - Sunday)?

• Approximately 0.64 account/s¹, 7.8 storage-slot/s
• Approximately 31.7B/s for accounts, 593B/s for storage ~= 54MB/day, 19.7GB/y

Catch: above growth is the pure useful state data

• Account trie weighs ~155.2B/acc, storage tries weigh ~142.3B/slot²
• Trie grows 99.3B/s for accounts, 1110B/s for storage ~= 104.5MB/day, 38.2GB/y³

¹+25% according to Etherscan; ²computer said so, maybe lied; ³disregarded log component

What does this all mean?

Ethereum (along with all forks) is on a – potential – death trajectory 😱

• Constant TPS ⇒ state growth ⇒ higher RAM ⇒ more IOPS ⇒ lower TPS ⇒ brick wall
• Mainnet can do a lot more TPS ⇒ brings the brick wall closer

But does the brick wall need to exist?

• EIP-4444: Bound Historical Data in Execution Clients
• EIP-4844: Shard Blob Transactions
• State rent or exponential costs

Thank you