Use upstream rocksdb #11248

dvdplm · 2019-11-11T18:45:49Z

…by way of paritytech/parity-common#257 by @ordian.

@ordian

…by way of paritytech/parity-common#257 by @ordian.

* master: Clarify what first_block `None` means (#11269) removed redundant VMType enum with one variant (#11266) Ensure jsonrpc threading settings are sane (#11267) Return Ok(None) when the registrar contract returns empty slice (#11257) Add a benchmark for snapshot::account::to_fat_rlps() (#11185) Fix misc compile warnings (#11258) simplify verification (#11249)

grbIzl · 2019-11-21T11:32:02Z

(Sorry for the stupid question) Do we have any metric or tests for it? I mean, do we understand, how will this affect client's performance? @ordian

dvdplm · 2019-11-21T12:11:21Z

Do we have any metric or tests for it?

Sort of. Both me and @ordian have been running block import tests over the last few days. Here are mine from last night, importing 5k blocks from approx mid-chain:

Master:

Benchmark #1: ./parities/parity-master --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     320.989 s ± 24.305 s    [User: 341.816 s, System: 41.915 s]
  Range (min … max):   302.979 s … 362.882 s    5 runs

Upstream rocksdb:

Benchmark #1: ./parities/parity-ao --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
   Time (mean ± σ):     283.643 s ± 10.213 s    [User: 282.681 s, System: 43.740 s]
   Range (min … max):   271.713 s … 296.393 s    5 runs

I think @ordian's results are similar.

* master: Upgrade trie-db to 0.16.0. (#11274)

ordian · 2019-11-21T15:32:48Z

@grbIzl I'm seeing 20% improvement on parity import when switching to this branch (tested on 100 and 5k recent blocks from mainnet at ~8.9m).

dvdplm · 2019-11-21T21:50:37Z

More benchmarks, now also comparing a master build from today (Nov 21, 2019), after the new trie landed (referred to below as "fast trie", but it's actually a little slower, cc @cheme @jimpo):

Upstream rocksdb

Benchmark #1: ./parities/parity-ao --mode=offline --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     284.806 s ±  7.885 s    [User: 283.799 s, System: 42.353 s]
  Range (min … max):   277.414 s … 296.277 s    5 runs

Upstream rocksdb + fast trie

Benchmark #1: ./parities/parity-ao-fast-trie --mode=offline --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     292.625 s ± 12.209 s    [User: 290.013 s, System: 42.446 s]
  Range (min … max):   282.579 s … 312.988 s    5 runs

Forked rocksdb (master, Nov 20)

Benchmark #1: ./parities/parity-master --mode=offline --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     321.283 s ±  8.235 s    [User: 340.501 s, System: 41.963 s]
  Range (min … max):   312.589 s … 333.771 s    5 runs

Forked rocksdb + fast trie (master, Nov 21)

Benchmark #1: ./parities/parity-master-fast-trie --mode=offline --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     323.682 s ±  7.001 s    [User: 341.980 s, System: 40.867 s]
  Range (min … max):   317.452 s … 334.709 s    5 runs

I also ran the same set of benchmarks with more verifier threads (8), but it has little impact.

Upstream rocksdb + 8 verif threads

Benchmark #1: ./parities/parity-ao --mode=offline --scale-verifiers --num-verifiers=8 --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     277.217 s ±  8.206 s    [User: 266.063 s, System: 41.339 s]
  Range (min … max):   268.896 s … 288.652 s    5 runs

Upstream rocksdb + fast trie + 8 verif threads

Benchmark #1: ./parities/parity-ao-fast-trie --mode=offline --scale-verifiers --num-verifiers=8 --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     288.779 s ±  8.539 s    [User: 274.010 s, System: 41.654 s]
  Range (min … max):   279.360 s … 301.326 s    5 runs

Forked rocksdb + 10 verif threads (master, Nov 20)

Benchmark #1: ./parities/parity-master --mode=offline --scale-verifiers --num-verifiers=8 --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     311.093 s ±  5.957 s    [User: 323.145 s, System: 40.403 s]
  Range (min … max):   301.396 s … 316.075 s    5 runs

Forked rocksdb + fast trie + 8 verif threads (master, Nov 21)

Benchmark #1: ./parities/parity-master-fast-trie --mode=offline --chain=eth --db-path=./current/bench-mainnet --pruning-history=12000 import ./blocks/blocks-5000.bin
  Time (mean ± σ):     322.392 s ±  7.124 s    [User: 341.089 s, System: 40.680 s]
  Range (min … max):   314.879 s … 333.632 s    5 runs

cheme · 2019-11-21T23:23:16Z

That's very interesting, trie db crate improvement are mainly on triedb iterator,
it is possible that it does not matter a lot for ethereum.
I suspect the loss come from decoding where we first build a nodeplan and then call 'build' function to go back to 'Node'. Before the pr 'Node' was directly produce from 'decode'. IIrc perf degrades on trie iter the same way at first but @jimpo fix that by avoiding 'build' calls it gets better. Probably it is the same here: we should avoid calling 'build' in triedbmut by having 'Node' replaced by some struct like '(data_slice, NodePlan)' but it is just a wild guess.

It could be rather simple to test: just put back the old 'decode' function in node_codec and see if it fix the perf (decode still exist in the trait but with a default implementation that call the 'build' function).

grbIzl · 2019-11-22T11:41:22Z

@ordian @dvdplm thanks for info! Looks promising. May be worth mentioning in PR's description also.

dvdplm · 2019-11-25T13:56:27Z

It could be rather simple to test: just put back the old 'decode' function in node_codec and see if it fix the perf (decode still exist in the trait but with a default implementation that call the 'build' function).

I did this and but my micro-benchmark did not show any meaningful regression, not sure if my benchmark is bad perhaps. :/

* master: [ethcore]: apply filter when `PendingSet::AlwaysQueue` in `ready_transactions_filtered` (#11227) Update lib.rs (#11286) Don't prune ancient state when instantiating a Client (#11270) fixed verify_uncles error type (#11276) ethcore: fix rlp deprecation warnings (#11280)

cheme · 2019-11-25T15:03:20Z

I did this and but my micro-benchmark did not show any meaningful regression, not sure if my benchmark is bad perhaps. :/

Thanks for checking 👍 , should be from somewhere else then :\ .

General-Beck · 2019-11-26T12:10:20Z

I think it will be necessary to test the build on the architectures aarch64, armv7, i386. And also on Windows and Mac platforms.
parity-clib library for android needs a separate test on the android device.

General-Beck · 2019-11-26T16:31:06Z

build test's https://gitlab.parity.io/General-Beck/parity-ethereum/pipelines/70139
CI updated, on Linux hosts installed clang-8, on windows clang-9

dvdplm · 2019-11-26T16:59:22Z

Linux hosts installed clang-8, on windows clang-9

Why not the same version on both? i.e. clang-9?

General-Beck · 2019-11-26T17:41:12Z

Why not the same version on both? i.e. clang-9?

During the discussion phase, I and @TriplEight do not mind using stable clang-9. We only need confirmation from the developers. and so that it does not interfere with other projects

dvdplm · 2019-11-26T18:13:54Z

confirmation from the developers. and so that it does not interfere with other projects

Hmm, how do you mean, "developers"? And I'm not sure which other projects share the same CI setup, can you elaborate?

dvdplm · 2019-12-01T19:30:14Z

Let me clarify. We've decided to move all our CI environments to clang-9.

Please let us know when the work is completed, thanks!

General-Beck · 2019-12-02T16:48:14Z

build passed

tomusdrw

Great! ~~Do we have some benchmarks?~~ NVM just saw in the comments.

parity/db/rocksdb/helpers.rs

parity/db/rocksdb/mod.rs

parity/db/rocksdb/helpers.rs

niklasad1

nice, LGTM

* master: add support for evan.network chains (#11289) Add benchmarks and tests for RlpNodeCodec decoding (#11287) upgrade vergen to 3.0 (#11293) interruptible test and build jobs (#11294) Istanbul HF on POA Sokol (#11282)

ordian · 2019-12-03T14:46:01Z

Do we have some benchmarks?

I've talked to the devops team about running the import benches on a daily basis or at least on-demand to track perf regressions, will let you know once it's done.

Use upstream rocksdb

a398ee5

…by way of paritytech/parity-common#257 by @ordian.

dvdplm self-assigned this Nov 11, 2019

dvdplm added A3-inprogress ⏳ Pull request is in progress. No review needed at this stage. F5-task 🎬 Generic label for things that just have to get done but are no issues in first place. M5-dependencies 🖇 Dependencies. labels Nov 11, 2019

dvdplm and others added 4 commits November 12, 2019 18:30

Hint at how parity db reset works in the error message

77462f2

migration-rocksdb: fix build

d1c94bb

Cargo.toml: use git dependency instead of path

19ac58d

Merge branch 'master' into dp/chore/use-upstream-rocksdb

aa87edd

* master: Upgrade trie-db to 0.16.0. (#11274)

ordian added 5 commits November 25, 2019 15:28

update to latest kvdb-rocksdb

30433a3

fix tests

c00c80c

saner default for light client

0e6b6b1

rename open_db to open_db_light

f0029cb

ordian added 2 commits November 26, 2019 15:57

update to latest kvdb-rocksdb

ebac589

moar update to latest kvdb-rocksdb

e6be5c7

ordian marked this pull request as ready for review November 28, 2019 15:17

ordian added this to the 2.7 milestone Nov 28, 2019

ordian requested review from niklasad1 and tomusdrw and removed request for niklasad1 December 3, 2019 11:27

tomusdrw approved these changes Dec 3, 2019

View reviewed changes

niklasad1 reviewed Dec 3, 2019

View reviewed changes

parity/db/rocksdb/helpers.rs Show resolved Hide resolved

niklasad1 reviewed Dec 3, 2019

View reviewed changes

parity/db/rocksdb/mod.rs Show resolved Hide resolved

ordian reviewed Dec 3, 2019

View reviewed changes

parity/db/rocksdb/helpers.rs Outdated Show resolved Hide resolved

Update parity/db/rocksdb/helpers.rs

2951584

niklasad1 reviewed Dec 3, 2019

View reviewed changes

parity/db/rocksdb/helpers.rs Show resolved Hide resolved

niklasad1 reviewed Dec 3, 2019

View reviewed changes

parity/db/rocksdb/helpers.rs Show resolved Hide resolved

niklasad1 approved these changes Dec 3, 2019

View reviewed changes

ordian added 2 commits December 3, 2019 15:42

add docs to memory_budget division

1002879

Merge branch 'master' into dp/chore/use-upstream-rocksdb

12fceb8

* master: add support for evan.network chains (#11289) Add benchmarks and tests for RlpNodeCodec decoding (#11287) upgrade vergen to 3.0 (#11293) interruptible test and build jobs (#11294) Istanbul HF on POA Sokol (#11282)

ordian added A8-looksgood 🦄 Pull request is reviewed well. and removed A0-pleasereview 🤓 Pull request needs code review. labels Dec 3, 2019

ordian merged commit f6c3d4c into master Dec 3, 2019

ordian deleted the dp/chore/use-upstream-rocksdb branch December 3, 2019 15:59

ordian mentioned this pull request Dec 10, 2019

Encapsulate access to the client for secret store #11232

Merged

7 tasks

ymonye mentioned this pull request Dec 11, 2019

CentOS 7.x (7.7.1908) - Parity v2.7.0 Installation: failed to run custom build command for `librocksdb-sys v6.2.4 #11326

Closed

ordian mentioned this pull request Jan 13, 2020

Wrong RocksDB.Options parameter #11233

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use upstream rocksdb #11248

Use upstream rocksdb #11248

dvdplm commented Nov 11, 2019

grbIzl commented Nov 21, 2019

dvdplm commented Nov 21, 2019

ordian commented Nov 21, 2019 •

edited

Loading

dvdplm commented Nov 21, 2019 •

edited

Loading

cheme commented Nov 21, 2019

grbIzl commented Nov 22, 2019

dvdplm commented Nov 25, 2019

cheme commented Nov 25, 2019

General-Beck commented Nov 26, 2019

General-Beck commented Nov 26, 2019

dvdplm commented Nov 26, 2019

General-Beck commented Nov 26, 2019 •

edited

Loading

dvdplm commented Nov 26, 2019

dvdplm commented Dec 1, 2019

General-Beck commented Dec 2, 2019

tomusdrw left a comment •

edited

Loading

niklasad1 left a comment

ordian commented Dec 3, 2019 •

edited

Loading

Use upstream rocksdb #11248

Use upstream rocksdb #11248

Conversation

dvdplm commented Nov 11, 2019

grbIzl commented Nov 21, 2019

dvdplm commented Nov 21, 2019

ordian commented Nov 21, 2019 • edited Loading

dvdplm commented Nov 21, 2019 • edited Loading

cheme commented Nov 21, 2019

grbIzl commented Nov 22, 2019

dvdplm commented Nov 25, 2019

cheme commented Nov 25, 2019

General-Beck commented Nov 26, 2019

General-Beck commented Nov 26, 2019

dvdplm commented Nov 26, 2019

General-Beck commented Nov 26, 2019 • edited Loading

dvdplm commented Nov 26, 2019

dvdplm commented Dec 1, 2019

General-Beck commented Dec 2, 2019

tomusdrw left a comment • edited Loading

Choose a reason for hiding this comment

niklasad1 left a comment

Choose a reason for hiding this comment

ordian commented Dec 3, 2019 • edited Loading

ordian commented Nov 21, 2019 •

edited

Loading

dvdplm commented Nov 21, 2019 •

edited

Loading

General-Beck commented Nov 26, 2019 •

edited

Loading

tomusdrw left a comment •

edited

Loading

ordian commented Dec 3, 2019 •

edited

Loading