Trying Out & Benchmarking Bcachefs On Linux 6.7

[ayumu]

Ask him if you believe to be so savvy about it. You can paste his replies here

#bcachefs:matrix.bcachefs.org

Oh please, let's instead try and isolate the author from time-wasting nonsensical distractions.

[User1239939]

Kent just replied in a post on his Patreon page to these Phoronix benchmarks.

[Mitch]

Kent has made a response to these benchmarks: https://www.patreon.com/posts/note-on-phoronix-92281382

EDIT: In this response, he gets different results. He also says:
"I suspect a lot of [the performance issues shown on Phoronix] might be from having CONFIG_BCACHEFS_DEBUG_TRANSACTIONS on by default."

[Mathias]

Kent has made a response to these benchmarks: https://www.patreon.com/posts/note-on-phoronix-92281382

Default mkfs options for all three filesystems.

I suspect a lot of this might be from having CONFIG_BCACHEFS_DEBUG_TRANSACTIONS on by default.

I don't understand. Did Michael enable DEBUG_TRANSACTIONS or is that a general reason why BCachefs may be slower?


To the people saying BCachefs is likely not optimized, Kent spent a lot of time optimizing the FS. He had an argument with the kernel devs, because he had a custom JIT-like thing, to unpack some data structure, because it would make some part 5% faster (making some pages WX, which the kernel team didn't like). IIRC they found a different approach that didn't impact performance much. Kent does absolutely care about performance. And I'm sure he's at least a bit surprised BCachefs gets destroyed in these benches.

[Mitch]

I don't understand. Did Michael enable DEBUG_TRANSACTIONS or is that a general reason why BCachefs may be slower?


To the people saying BCachefs is likely not optimized, Kent spent a lot of time optimizing the FS. He had an argument with the kernel devs, because he had a custom JIT-like thing, to unpack some data structure, because it would make some part 5% faster (making some pages WX, which the kernel team didn't like). IIRC they found a different approach that didn't impact performance much. Kent does absolutely care about performance. And I'm sure he's at least a bit surprised BCachefs gets destroyed in these benches.

I think Kent is saying that BCacheFS itself leaves that option on by default. Possibly because it's still in Linux Next and 6.7 RC. That's how I read it. I don't think Michael would intentionally turn on a debug feature. He'd know that could only hurt benches.

[Malsabku]

I hope the poor performance is not due to mistakes in the architecture of the file system, but can be improved through optimisations.

[Hans Bull]

So was DEBUG_TRANSACTIONS on nor not?

[timofonic]

I don't understand. Did Michael enable DEBUG_TRANSACTIONS or is that a general reason why BCachefs may be slower?


To the people saying BCachefs is likely not optimized, Kent spent a lot of time optimizing the FS. He had an argument with the kernel devs, because he had a custom JIT-like thing, to unpack some data structure, because it would make some part 5% faster (making some pages WX, which the kernel team didn't like). IIRC they found a different approach that didn't impact performance much. Kent does absolutely care about performance. And I'm sure he's at least a bit surprised BCachefs gets destroyed in these benches.

According to Kent Overstreet in Bcachefs matrix chatroom:

Something is way off with phoronix's setup. In my testing Bcachefs comes out faster than btrfs and both are quite a bit faster than Phoronix's numbers.

Oh, i think a lot of why the phoronix numbers were so low. It is because CONFIG_BCACHEFS_DEBUG_TRANSACTIONS is on by default. It is a brilliant idea for how we can make all our btree_trans objects accessable for debugfs without them being on a single contended list.

Michael

[ptr1337]

So was DEBUG_TRANSACTIONS on nor not?

Yes, it was. The patch does enable it as default, as soon bcachefs gets enabled:

+config BCACHEFS_DEBUG_TRANSACTIONS
+ bool "bcachefs runtime info"
+ depends on BCACHEFS_FS
+ default y
+ help
+ This makes the list of running btree transactions available in debugfs.
+
+ This is a highly useful debugging feature but does add a small amount of overhead.​

But even the description of this debug features, that it adds a small amount of overhead would be just kinda wrong, more likely a insane overhead.

[timofonic]

I don't understand. Did Michael enable DEBUG_TRANSACTIONS or is that a general reason why BCachefs may be slower?


To the people saying BCachefs is likely not optimized, Kent spent a lot of time optimizing the FS. He had an argument with the kernel devs, because he had a custom JIT-like thing, to unpack some data structure, because it would make some part 5% faster (making some pages WX, which the kernel team didn't like). IIRC they found a different approach that didn't impact performance much. Kent does absolutely care about performance. And I'm sure he's at least a bit surprised BCachefs gets destroyed in these benches.

Despite of that, he still didn't focus on performance optimization yet.

He said in chatroom:

no, nothing i haven't sent out yet includes performance improvements