Hardware requirements

This page holds the disk, memory and time figures for running a node. Each one was measured on the machine described below, with the client’s default settings unless its row says otherwise, and each row records the date and the block it was measured at. The chain grows with every block, so a figure describes what that date needed.

The reference machine used for every figure below:

• CPU: Intel Core i7-10710U, 6 cores, 12 threads
• RAM: 31.1 GiB
• Disk: NVMe SSD, formatted btrfs
• Default settings: snap sync, --gcmode full, the hash state scheme for a new database, and --cache 1024

Ethereum Classic¶

Mordor¶

Stopping¶

How long a clean stop took, from the interrupt to Blockchain stopped in the log, and to the process exiting.

How each figure was measured¶

Time to sync runs from the node’s first log line to Snap sync complete, auto disabling, with both timestamps read from the log the node wrote to standard error:

1
2

$ grep -m1 -F 'Starting Core-Geth' geth.log
$ grep -m1 -F 'Snap sync complete, auto disabling' geth.log

Disk is du -sb on each directory, read within two minutes of that line. -b reports the apparent size in bytes, which the tables convert to GiB. Give du one directory per command: it counts each directory once per command, so nested directories passed together leave one of them out of a total.

1
2
3

$ du -sb <datadir>/geth/chaindata
$ du -sb <datadir>/geth/chaindata/ancient
$ du -sb <datadir>

On Ethereum Classic the node was still generating its state snapshot when the disk was read, and that snapshot is written into chaindata as it grows. On Mordor, generation had finished.

Peak memory is the process’s peak resident set size as the kernel reports it, read after the sync finished:

1

$ grep VmHWM /proc/<pid>/status

Load average is the first field of /proc/loadavg, sampled once a minute. It includes other work running on the machine at the same time.

Stop times run from Got interrupt, shutting down... to Blockchain stopped in the log. Process exit was checked once a second, so those times are to the nearest second.

Memory and --cache¶

--cache is the memory, in MiB, that the node gives its internal caches. The peak memory figures above use the default, 1024, which the node divides among --cache.database, --cache.trie, --cache.gc and --cache.snapshot. A default start logs the first three:

1
2

INFO [09-13|08:13:49.777] Allocated trie memory caches             clean=154.00MiB dirty=256.00MiB
INFO [09-13|08:13:49.777] Allocated cache and file handles         database=<datadir>/geth/chaindata cache=512.00MiB handles=524,288

The node lowers a --cache larger than one third of the machine’s memory, and logs the change. Asked for 20000 on the reference machine, it used 10601:

1

WARN [09-13|09:55:06.321] Sanitizing cache to Go's GC limits       provided=20000 updated=10601

On a 32-bit platform it counts no more than 2 GiB of memory when it sets that limit.

The ancient store on another disk¶

--datadir.ancient puts the ancient store, the older headers, bodies and receipts that otherwise live in chaindata/ancient, in a directory you choose. The tables above show how much of chaindata it holds. History on another disk moves the store of a synced node.

Pass the same --datadir.ancient on every start. Started without it, the node looks for the store inside chaindata instead. A node whose first blocks have already moved into its ancient store then refuses to start, with ancient chain segments already extracted, please set --datadir.ancient to the correct path. A node that has not started syncing creates a new, empty store there, and logs no error.

Running out of disk¶

The node checks the free space in <datadir>/geth every 30 seconds. Below twice its threshold, it warns:

1

WARN [09-13|09:55:10.945] Disk space is running low. Geth will shutdown if disk space runs below critical level. available=55.12GiB critical_level=35.80GiB path=<datadir>/geth

Below the threshold, it shuts itself down cleanly, as it would on an interrupt:

1
2

ERROR[09-13|09:55:10.091] Low disk space. Gracefully shutting down Geth to prevent database corruption. available=55.12GiB path=<datadir>/geth
INFO [09-13|09:55:10.091] Got interrupt, shutting down...

The threshold is twice the share of --cache given to --cache.gc, which is 512 MiB at the defaults. --datadir.minfreedisk sets it in MiB, and 0 turns the check off.

Small machines¶

Nothing on this page was measured on a smaller machine. Two limits matter on one, and both have their own section: key derivation on a small machine, and which ARM archive to install.