Will it support TRIM operation?

[Kaydannik]

Hello.
I have Samsung 840 Pro 512GB.
During performance test with filesystem on it - it give me ~ 400 mb/s writes, and up to 15 000 iops.

But if i add it like a flashcache - speed is falling rapidly. Also writing/reading speed is not stable.
I already excluded all variants with dirty_thresh_pct and fallow_delay, etc..

Same performance issues i have if i use SSD with ext4 without discard option.

So, is any news/opportunities to support TRIM in flashcache ?
Thanks

[llcooljayr]

I tested Samsung 840 Evo 512GB with flashcache in raid setup and saw the same performance degradation. You might want to check with Samsung because they might have provided a firmware update for the Pro line since they provided a firmware update for the Evo which fixed an algorithm problem that address the dropped in performance. Also I don't think trim is supported by flashcache. In order to perform trim, the operating system, filesystem, ssd, and flashcache has to support it.

[mohans]

We have changes for TRIM support that will be committed shortly.

---

From: Kaydannik notifications@github.com
To: facebook/flashcache flashcache@noreply.github.com
Sent: Friday, November 7, 2014 5:44 AM
Subject: [flashcache] Will it support TRIM operation? (#192)

Hello.
I have Samsung 840 Pro 512GB.
During performance test with filesystem on it - it give me ~ 400 mb/s writes, and up to 15 000 iops.
But if i add it like a flashcache - speed is falling rapidly. Also writing/reading speed is not stable.
I already excluded all variants with dirty_thresh_pct and fallow_delay, etc..
Same performance issues i have if i use SSD with ext4 without discard option.
So, is any news/opportunities to support TRIM in flashcache ?
Thanks
—
Reply to this email directly or view it on GitHub.

[Kaydannik]

Awesome! champagne is prepared!

[rferris]

I'm so excite!

[Kaydannik]

@mohans Hi. Any news with TRIM ? Is project is in development ?

[mohans]

Unfortunately, all new development on flashcache is on hold now (except for any critical fixes). Too busy working on other things :(, will revisit the trim support later this year as time permits.

[ghost]

Hi,
I'm looking at the code trying to find place for TRIM.
And I'm curious if putting ioctl(device, BLKDISCARD, &range); in flashcache_subr.c:flashcache_invalid_remove() is enough?

[ghost]

Sorry, my mistake, not flashcache_invalid_remove(), but flashcache_invalid_insert().

[ghost]

https://github.com/sphdev/flashcache/commit/76b44f1c62f902f51261da44cb98398df868175c
Trying to add TRIM support.
Still not sure if it's enough.

[ghost]

Another interesting place for TRIM could be flashcache_hash_remove()
Just thinking how many places like this exist and which one is the best for TRIM command. Any ideas?
https://github.com/sphdev/flashcache/commit/94ae1cd614b12c371d2cc5404a7892662f8ccc92

[ghost]

I've been trying to implement support for FITRIM.
I'm using Debian, my partition is formated as XFS. So I've started with checking, why FITRIM is not passed.
In flashcache_conf.c in function flashcache_ctr() I've added two lines:
ti->discards_supported = 1;
ti->num_discard_requests = SOMEVALUE; // I have no idea what should be there
With those 2 lines FITRIM in XFS is passed. But...
The problem is kernel panic that occurred during flashcache_map() in flashcache_main.c.
This function is calling flashcache_do_block_checks().
Inside we have:
VERIFY(to_sector(bio->bi_size) <= dmc->block_size);
And the problem is to_sector(bio->bi_size), because it's too high. In my situation it's 2280560 and dmc->block_size is 8.
Does anyone have any idea how to solve this problem?

My test looks like this:
#flashcache_load /dev/sdb
#mount /dev/mapper/fcache /mnt
#fstrim /mnt
(kernel panic)

[mohans]

You need a lot more changes for TRIM that those changes in _conf.c.
TRIM is still a work in progress. We have patches (not committed) that discard TRIM'ed blocks from the cache, but those patches do not yet send the TRIMs down to the cache drive.
We have been pulled to work on other things, so don't know when those patches will be committed.
mohan
From: sphdev notifications@github.com
To: facebook/flashcache flashcache@noreply.github.com
Cc: Mohan Srinivasan mohan_srinivasan@yahoo.com
Sent: Monday, July 13, 2015 2:00 AM
Subject: Re: [flashcache] Will it support TRIM operation? (#192)

I've been trying to implement support for FITRIM.
I'm using Debian, my partition is formated as XFS. So I've started with checking, why FITRIM is not passed.
In flashcache_conf.c in function flashcache_ctr() I've added two lines:
ti->discards_supported = 1;
ti->num_discard_requests = SOMEVALUE; // I have no idea what should be there
With those 2 lines FITRIM in XFS is passed. But...
The problem is kernel panic that occurred during flashcache_map() in flashcache_main.c.
This function is calling flashcache_do_block_checks().
Inside we have:
VERIFY(to_sector(bio->bi_size) <= dmc->block_size);
And the problem is to_sector(bio->bi_size), because it's too high. In my situation it's 2280560 and dmc->block_size is 8.
Does anyone have any idea how to solve this problem?My test looks like this:
#flashcache_load /dev/sdb
#mount /dev/mapper/fcache /mnt
#fstrim /mnt
(kernel panic)—
Reply to this email directly or view it on GitHub.

[ghost]

Is there any way to see those patches?