sync issue for big databases (archive nodes)

[kogeler]

role: full(archive)
binary: docker pull parity/polkadot:v0.9.42
instance: GCP - t2d-standard-4
disk: GCP - SSD persistent disk
OS: Container-Optimized OS from Google
kernel: 5.10.162+
CLI flags:

--name=${POD_NAME} \
--base-path=/chain-data \
--keystore-path=/keystore \
--chain=${CHAIN} \
--database=paritydb \
--pruning=archive \
--prometheus-external \
--prometheus-port 9615 \
--unsafe-rpc-external \
--unsafe-ws-external \
--rpc-cors=all \
--in-peers 75 \
--out-peers 25 \
--public-addr=/ip4/${EXTERNAL_IP}/tcp/${RELAY_CHAIN_P2P_PORT} \
 --listen-addr=/ip4/0.0.0.0/tcp/30333 \

I'm trying to sync backup nodes from scratch. I have 8 nodes (Kusama, Polkadot, archive, prune, rocksdb, paritydb).

I use the same instances, regions, and CLI flags.
All nodes have 100 peers (in 75/out 25).

2 archive rocksdb nodes (Kusama, Pollkadot) synced in a couple of days.
But 2 archive paritydb nodes (Kusama, Pollkadot) have been syncing for 1,5 weeks. At some point (around 15M blocks), the sync rate decreased quickly. Now it is less than 0 blocks/second. Restars don't help.
It looks like an issue of paritydb.

The current state is:
Kusama - target=#18852634 (100 peers), best: #15387848 (0x117e…c8a0), finalized #15387648 (0x065b…4684), ⬇ 705.8kiB/s ⬆ 461.9kiB/s
Polkadot - target=#16463661 (100 peers), best: #15045441 (0x9fc3…0db0), finalized #15045402 (0xfef9…401b), ⬇ 134.9kiB/s ⬆ 125.2kiB/s

The disk sub-system is overloaded: 15k iops and 100MB/s by reading.

[kogeler]

Kusama DB (1.1TB): https://storage.googleapis.com/debug-releases/debug/debug-paritydb.tar

[kogeler]

ref: #212

[arkpar]

@kogeler Woud it be possible to get SSH access to the machine as well?

[kogeler]

@arkpar I think it is not possible because it's just a pod in a k8s cluster. Yes, it uses a dedicated k8s node, but it isn't trivial to connect to the runtime environment. I shutdown the pod and made a snapshot of the GCP disk to upload the copy of the DB.

[arkpar]

Once the index file stops fitting in memory, hash index search becomes a major bottleneck. There will be some improvenets in the next parity-db release, but ultimately this will be resolve with #199, which will remove index lookups for trie nodes.

[arkpar]

@kogeler Could you then upload a copy of rocksdb database as well? For some reference testing.

[kogeler]

@arkpar Do you need a full-synced archive copy of rocksdb?

[arkpar]

@kogeler Fully synced or around the same block as the parity-db snapshot (15m -ish). If it is too much trouble, I can probably sync one myself, even though it will take a few days.

[kogeler]

@arkpar You can download our public periodic DB snapshots using the manual

[kogeler]

@arkpar Are there any updates?

[arkpar]

We are working on a major feature (#199) that will resolve this. It will take a few weaks to land in substrate/polkadot.

[kogeler]

@arkpar Are there any updates about this issue?