core/state: improve update times #27
Conversation
|
"master" = snapshot-5 Currently doing a fast-sync, so we can later compare the times for account and storage updates. |
|
restarted |
holiman
force-pushed
the
improve_updates
branch
2 times, most recently
from
636bf14 to
89416e1
Compare
|
Trying out a wild idea. Updated the machines: |
|
The change in Before the interruption, the This pr-machine is at
Next up: I'll try the same trick for storage. |
|
Here's the gist of it. First of all, scrap the existing precacher. Background:When we execute block IdeaSo, during finalize, we also send each So the snapshotter allows us to parallelize the trie-reads instead of doing it serially. At the end of the block, when we finally reach the update-phase, the trie caches will be prepopulated already, and the update will be quick. |
|
The docker stop timeout of 60s might be insufficient, one of them got corrupted and I had to resync. After both are in sync again, and snapshot generation complete, However, here's the The corresponding one for this pr is a couple of milliseconds better:
The new prefetcher seems ot be loading about 50 entries per second, which is ~700 items per block, which seems a bit on the low side -- we're probably missing a lot of things due to async sends.
|
|
After sync-up, the The Also, looking at the clean cache hits, one can see that
|
|
Oh, this is a nice graph aswell: |
holiman
changed the title
|
I am now testing to swap the nodes around, so this PR runs on |
|
Ok, here are some stats from switching them The upper one is first The lower one was this PR, then TLDR; this PR makes account/storage updates 27-39 % faster. |
|
Interestingly, the |
karalabe
force-pushed
the
snapshot-5
branch
6 times, most recently
from
586c663 to
06d4470
Compare
This makes eth_call and eth_estimateGas use the zero address as sender when the "from" parameter is not supplied. Co-authored-by: Felix Lange <[email protected]>
* les: separate peer into clientPeer and serverPeer * les: address comments
This was missing because I forgot to wrap it when bind.CallOpts.From as added.
|
The measurements of this PR vs without is not comparable because you modified the metric measuring account/storage updates. That change was never upstreamed into the original PR. |
typo in func name in the comment
…m#20746) Includes difficulty tests for EIP2384 aka MuirGlacier.
This is supposed to fix the occasional failures in TestCancel* on Travis CI.
eth: fix transaction announce/broadcast goroutine leak
go.mod: update golang.org/x/crypto to fix a Go 1.14 race rejection
This revision of go-duktype fixes the following warning
```
duk_logging.c: In function ‘duk__logger_prototype_log_shared’:
duk_logging.c:184:64: warning: ‘Z’ directive writing 1 byte into a region of size between 0 and 9 [-Wformat-overflow=]
184 | sprintf((char *) date_buf, "%04d-%02d-%02dT%02d:%02d:%02d.%03dZ",
| ^
In file included from /usr/include/stdio.h:867,
from duk_logging.c:5:
/usr/include/x86_64-linux-gnu/bits/stdio2.h:36:10: note: ‘__builtin___sprintf_chk’ output between 25 and 85 bytes into a destination of size 32
36 | return __builtin___sprintf_chk (__s, __USE_FORTIFY_LEVEL - 1,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
37 | __bos (__s), __fmt, __va_arg_pack ());
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
```
Fixes: Condition is always 'false' because 'err' is always 'nil'
This PR fixes issues in TableDatabase. TableDatabase is a wrapper of underlying ethdb.Database with an additional prefix. The prefix is applied to all entries it maintains. However when we try to retrieve entries from it we don't handle the key properly. In theory the prefix should be truncated and only user key is returned. But we don't do it in some cases, e.g. the iterator and batch replayer created from it. So this PR is the fix to these issues.
eth: when triggering a sync, check the head header TD, not block
core/rawdb: fix freezer table test error check
internal/web3ext: fix clique console apis to work on missing arguments
These tests occasionally fail on Travis.
…0776) This just prevents a false negative ERROR warning when, for some unknown reason, a user attempts to turn on the module rpc even though it's already going to be on.
Dynamic state snapshots
holiman
force-pushed
the
improve_updates
branch
from
56c37fd to
91b0a77
Compare
eth/downloader: restart the downloader after completion on new head
Work in progress to improve the
account_update/storage_update.Background: the chart below shows switching from snapshot-5 to
master, and how theaccount_updateandstorage_updateeats up the gains we get from the account/storage-read operations.