On 6/10/21 1:06 PM, Matthew Wilcox wrote:
On Thu, Jun 10, 2021 at 12:22:40PM -0400, Ric Wheeler wrote:
On 6/9/21 5:32 PM, Jaegeuk Kim wrote:
On Wed, Jun 9, 2021 at 11:47 AM Bart Van Assche <bvanassche@xxxxxxx
<mailto:bvanassche@xxxxxxx>> wrote:
On 6/9/21 11:30 AM, Matthew Wilcox wrote:
> maybe you should read the paper.
>
> " Thiscomparison demonstrates that using F2FS, a flash-friendly file
> sys-tem, does not mitigate the wear-out problem, except inasmuch asit
> inadvertently rate limitsallI/O to the device"
Do you agree with that statement based on your insight? At least to me, that
paper is missing the fundamental GC problem which was supposed to be
evaluated by real workloads instead of using a simple benchmark generating
4KB random writes only. And, they had to investigate more details in FTL/IO
patterns including UNMAP and LBA alignment between host and storage, which
all affect WAF. Based on that, the point of the zoned device is quite promising
to me, since it can address LBA alignment entirely and give a way that host
SW stack can control QoS.
Just a note, using a pretty simple and optimal streaming write pattern, I
have been able to burn out emmc parts in a little over a week.
My test case creating a 1GB file (filled with random data just in case the
device was looking for zero blocks to ignore) and then do a loop to cp and
sync that file until the emmc device life time was shown as exhausted.
This was a clean, best case sequential write so this is not just an issue
with small, random writes.
How many LBAs were you using? My mental model of a FTL (which may
be out of date) is that it's essentially a log-structured filesystem.
When there are insufficient empty erase-blocks available, the device
finds a suitable victim erase-block, copies all the still-live LBAs into
an active erase-block, updates the FTL and erases the erase-block.
So the key is making sure that LBAs are reused as much as possible.
Short of modifying a filesystem to make this happen, I force it by
short-stroking my SSD. We can model it statistically, but intuitively,
if there are more "live" LBAs, the higher the write amplification and
wear on the drive will be because the victim erase-blocks will have
more live LBAs to migrate.
This is why the paper intrigued me; it seemed like they were rewriting
a 100MB file in place. That _shouldn't_ cause ridiculous wear, unless
the emmc device was otherwise almost full.
During the test run, I did not look at which LBA's that got written to over the
couple of weeks.
Roughly, I tried to make sure that the file system ranged in fullness from 50%
to 75% (did not let it get too close to full).
Any vendor (especially on the low end parts) might do something really
primitive, but the hope I would have is similar to what you describe - if there
is sufficient free space, the firmware should be able to wear level across all
of the cells in the device. Overwriting in place or writing (and then
freeing/discarded) each LBA *should* be roughly equivalent. Free space being
defined as LBA's that are not known to the device as those without valid,
un-discarded data.
Also important to write enough to flush through any possible DRAM/SRAM like
cache a device might have that could absorb tiny writes.
The parts I played with ranged from what seemed to be roughly 3x write
amplification for the workload I ran down to more like 1.3x write amplification
(measured super coarsely as app level IO dispatched so all metadata, etc counted
as "WA" in my coarse look). Just trying to figure out for a given IO/fs stack,
how specific devices handle the user workload.
Regards,
Ric
