ethdb: add pebble for experiment #21575
Conversation
rjl493456442
requested review from
fjl,
holiman,
karalabe and
renaynay
as code owners
rjl493456442
force-pushed
the
exp-pebble
branch
2 times, most recently
from
65cace4 to
1d689c5
Compare
|
I rebased this and ran it on So the pebble is only Some charts (this PR is yellow)
This PR on the right-hand side (and in this chart, this PR seems to be slower) |
|
For the last 12 hours, the pebble machine has been using ~1.5GB less memory, so the comparisons aren't quite fair. We should give it a bit more memory. |
That's 3%. Don't count the chain which lives outside of the database. Curious why out previous benchmarks got 20%+. Could you start a full sync perhaps? |
|
restarting this as a full sync now: https://geth-bench.ethdevops.io/d/Jpk-Be5Wk/dual-geth?orgId=1&from=1606125272921&to=now |
|
It's been running a little more than one day now, and is consistently ~2GB lower in memory consumption
However, it's also about 1hours behind on the blocks. Size-wise Leveldb: So it's quite a hefty difference: |
|
Another update. Pebble is yellow, leveldb is green. They're both roughly in the same area of blocks But leveldb is ~6h ahead:
Pebbles consistently uses ~2Gb less memory: Pebbble has done a lot less disk reads:
But roughly the same amount of writes:
Disk-usage-wise, pebble has used All in all, the benchmark doesn't seem very fair, since pebble is using 2G less memory. |
| func (db *Database) NewBatch() ethdb.Batch { | ||
| return &batch{ | ||
| b: db.db.NewBatch(), | ||
| } | ||
| } |
There was a problem hiding this comment.
Pebble batches should be Closed after use -- they are pooled internally, and Close put it back into the pool. This requires changes to our batch API.
| // Pebble has a single combined cache area and the write | ||
| // buffers are taken from this too. Assign all available | ||
| // memory allowance for cache. | ||
| Cache: pebble.NewCache(int64(cache * 1024 * 1024)), |
There was a problem hiding this comment.
From pebble docs
// The cache is created with a
// reference count of 1. Each DB it is associated with adds a reference, so the
// creator of the cache should usually release their reference after the DB is
// created.
//
// c := pebble.NewCache(...)
// defer c.Unref()
// d, err := pebble.Open(pebble.Options{Cache: c})
|
On a new sync, where I gave pebble a bit more cache, and had it run on |
|
We should probably try this another time. |
Heh, I was looking at this PR only yesterday, thinking the exact same thing. Apart from just "using" pebble, we could also use their delete-range operations. Currently, we do range-deletion "manually", and don't have in in the database interface (we just iterate and delete). We do make use of it pretty much though, particularly during snap sync. |
ethdb/pebble: disable sync-style db write ethdb/pebble: update configs ethdb/pebble: fixes and tests ethdb/pebble: update configs
|
Rebased on master, updated to latest pebble |
|
Running again, this PR on |
|
New version: #24615 |
No description provided.