Re: Desperate with 870 QVO and ZFS

From: Stefan Esser <>
Date: Fri, 08 Apr 2022 11:14:24 UTC
Am 07.04.22 um 14:30 schrieb
> El 2022-04-06 23:49, Stefan Esser escribió:
>>> El 2022-04-06 17:43, Stefan Esser escribió:
>>>     Am 06.04.22 um 16:36 schrieb
>>>         Hi Rainer!
>>>         Thank you so much for your help :) :)
>>>         Well I assume they are in a datacenter and should not be a power
>>>         outage....
>>>         About dataset size... yes... our ones are big... they can be 3-4 TB
>>>         easily each
>>>         dataset.....
>>>         We bought them, because as they are for mailboxes and mailboxes
>>>         grow and
>>>         grow.... for having space for hosting them...
>>>     Which mailbox format (e.g. mbox, maildir, ...) do you use?
>>>     *I'm running Cyrus imap so sort of Maildir... too many little files
>>>     normally..... Sometimes directories with tons of little files....*
>> Assuming that many mails are much smaller than the erase block size of the
>> SSD, this may cause issues. (You may know the following ...)
>> For example, if you have message sizes of 8 KB and an erase block size of 64
>> KB (just guessing), then 8 mails will be in an erase block. If half the
>> mails are deleted, then the erase block will still occupy 64 KB, but only
>> hold 32 KB of useful data (and the SSD will only be aware of this fact if
>> TRIM has signaled which data is no longer relevant). The SSD will copy
>> several partially filled erase blocks together in a smaller number of free
>> blocks, which then are fully utilized. Later deletions will repeat this
>> game, and your data will be copied multiple times until it has aged (and the
>> user is less likely to delete further messages). This leads to "write
>> amplification" - data is internally moved around and thus written multiple
>> times.
>> *Stefan!! you are nice!! I think this could explain all our problem. So, why
>> we are having the most randomness in our performance degradation and that
>> does not necessarily has to match with the most io peak hours... That I
>> could cause that performance degradation just by deleting a couple of huge
>> (perhaps 200.000 mails) mail folders in a middle traffic hour time!!*
Yes, if deleting large amounts of data triggers performance issues (and the
disk does not have a deficient TRIM implementation), then the issue is likely
to be due to internal garbage collections colliding with other operations.
>> *The problem is that by what I know, erase block size of an SSD disk is
>> something fixed in the disk firmware. I don't really know if perhaps it
>> could be modified with Samsung magician or those kind of tool of Samsung....
>> else I don't really see the manner of improving it... because apart from
>> that, you are deleting a file in raidz-2 array... no just in a disk... I
>> assume aligning chunk size, with record size and with the "secret" erase
>> size of the ssd, perhaps could be slightly compensated?.*
The erase block size is a fixed hardware feature of each flash chip. There is a
block size for writes (e.g. 8 KB) and many such blocks are combined in one
erase block (of e.g. 64 KB, probably larger in todays SSDs), they can only be
returned to the free block pool all together. And if some of these writable
blocks hold live data, they must be preserved by collecting them in newly
allocated free blocks.

An example of what might happen, showing a simplified layout of files 1, 2, 3
(with writable blocks 1a, 1b, ..., 2a, 2b, ... and "--" for stale data of
deleted files, ".." for erased/writable flash blocks) in an SSD might be:

erase block 1: |1a|1b|--|--|2a|--|--|3a|

erase block 2; |--|--|--|2b|--|--|--|1c|

erase block 3; |2c|1d|3b|3c|--|--|--|--|

erase block 4; |..|..|..|..|..|..|..|..|

This is just a random example how data could be laid out on the physical
storage array. It is assumed that the 3 erase blocks once were completely occupied

In this example, 10 of 32 writable blocks are occupied, and only one free erase
block exists.

This situation must not persist, since the SSD needs more empty erase blocks.
10/32 of the capacity is used for data, but 3/4 of the blocks are occupied and
not immediately available for new data.

The garbage collection might combine erase blocks 1 and 3 into a currently free
one, e.g. erase block 4:

erase block 1; |..|..|..|..|..|..|..|..|

erase block 2; |--|--|--|2b|--|--|--|1c|

erase block 3; |..|..|..|..|..|..|..|..|

erase block 4: |1a|1b|2a|3a|2c|1d|3b|3c|

Now only 2/4 of the capacity is not available for new data (which is still a
lot more than 10/32, but better than before).

Now assume file 2 is deleted:

erase block 1; |..|..|..|..|..|..|..|..|

erase block 2; |--|--|--|--|--|--|--|1c|

erase block 3; |..|..|..|..|..|..|..|..|

erase block 4: |1a|1b|--|3a|--|1d|3b|3c|

There is now a new sparsely used erase block 4, and it will soon need to be
garbage collected, too - in fact it could be combined with the live data from
erase block 2, but this may be delayed until there is demand for more erased
blocks (since e.g. file 1 or 3 might also have been deleted by then).

The garbage collection does not know which data blocks belong to which file,
and therefore it cannot collect the data belonging to a file into a single
erase block. Blocks are allocated as data comes in (as long as enough SLC cells
are available in this area, else directly in QLC cells). Your many parallel
updates will cause fractions of each larger file to be spread out over many
erase blocks.

As you can see, a single file that is deleted may affect many erase blocks, and
you have to take redundancy into consideration, which will multiply the effect
by a factor of up to 3 for small files (one ZFS allocation block). And
remember: deleting a message in mdir format will free the data blocks, but will
also remove the directory entry, causing additional meta-data writes (again
multiplied by the raid redundancy).

A consumer SSD would normally see only very few parallel writes, and sequential
writes of full files will have a high chance to put the data of each file
contiguously in the minimum number of erase blocks, allowing to free multiple
complete erase blocks when such a file is deleted and thus obviating the need
for many garbage collection copies (that occur if data from several independent
files is in one erase block).

Actual SSDs have many more cells than advertised. Some 10% to 20% may be kept
as a reserve for aging blocks that e.g. may have failed kind of a
"read-after-write test" (implemented in the write function, which adds charges
to the cells until they return the correct read-outs).

BTW: Having an ashift value that is lower than the internal write block size
may also lead to higher write amplification values, but a large ashift may lead
to more wasted capacity, which may become an issue if typical file length are
much smaller than the allocation granularity that results from the ashift value.

>> Larger mails are less of an issue since they span multiple erase blocks,
>> which will be completely freed when such a message is deleted.
>> *I see I see Stefan...*
>> Samsung has a lot of experience and generally good strategies to deal with
>> such a situation, but SSDs specified for use in storage systems might be
>> much better suited for that kind of usage profile.
>> *Yes... and the disks for our purpose... perhaps weren't QVOs....*
You should have got (much more expensive) server grade SSDs, IMHO.

But even 4 * 2 TB QVO (or better EVO) drives per each 8 TB QVO drive would
result in better performance (but would need a lot of extra SATA ports).

In fact, I'm not sure whether rotating media and a reasonable L2ARC consisting
of a fast M.2 SSD plus a mirror of small SSDs for a LOG device would not be a
better match for your use case. Reading the L2ARC would be very fast, writes
would be purely sequential and relatively slow, you could choose a suitable
L2ARC strategy (caching of file data vs. meta data), and the LOG device would
support fast fsync() operations required for reliable mail systems (which
confirm data is on stable storage before acknowledging the reception to the

>>>         We knew they had some speed issues, but those speed issues, we
>>>         thought (as
>>>         Samsung explains in the QVO site) they started after exceeding the
>>>         speeding
>>>         buffer this disks have. We though that meanwhile you didn't exceed it's
>>>         capacity (the capacity of the speeding buffer) no speed problem
>>>         arises. Perhaps
>>>         we were wrong?.
>>>     These drives are meant for small loads in a typical PC use case,
>>>     i.e. some installations of software in the few GB range, else only
>>>     files of a few MB being written, perhaps an import of media files
>>>     that range from tens to a few hundred MB at a time, but less often
>>>     than once a day.
>>>     *We move, you know... lots of little files... and lot's of different
>>>     concurrent modifications by 1500-2000 concurrent imap connections we
>>>     have...*
>> I do not expect the read load to be a problem (except possibly when the SSD
>> is moving data from SLC to QLC blocks, but even then reads will get
>> priority). But writes and trims might very well overwhelm the SSD,
>> especially when its getting full. Keeping a part of the SSD unused (excluded
>> from the partitions created) will lead to a large pool of unused blocks.
>> This will reduce the write amplification - there are many free blocks in the
>> "unpartitioned part" of the SSD, and thus there is less urgency to compact
>> partially filled blocks. (E.g. if you include only 3/4 of the SSD capacity
>> in a partition used for the ZPOOL, then 1/4 of each erase block could be
>> free due to deletions/TRIM without any compactions required to hold all this
>> data.)
>> Keeping a significant percentage of the SSD unallocated is a good strategy
>> to improve its performance and resilience.
>> *Well, we have allocated all the disk space... but not used... just
>> allocated.... you know... we do a zpool create with the whole disks.....*
I think the only chance for a solution that does not require new hardware is to
make sure, only some 80% of the SSDs are used (i.e. allocate only 80% for ZFS,
leave 20% unallocated). This will significantly reduce the rate of garbage
collections and thus reduce the load they cause.

I'd use a fast encryption algorithm (zstd - choose a level that does not
overwhelm the CPU, there are benchmark results for ZFS with zstd, and I found
zstd-2 to be best for my use case). This will more than make up for the space
you left unallocated on the SSDs.

A different mail box format might help, too - I'm happy with dovecot's mdbox
format, which is as fast but much more efficient than mdir.

>>>     As the SSD fills, the space available for the single level write
>>>     cache gets smaller
>>>     *The single level write cache is the cache these ssd drivers have, for
>>>     compensating the speed issues they have due to using qlc memory?. Do
>>>     you refer to that?. Sorry I don't understand well this paragraph.*
>> Yes, the SSD is specified to hold e.g. 1 TB at 4 bits per cell. The SLC
>> cache has only 1 bit per cell, thus a 6 GB SLC cache needs as many cells as
>> 24 GB of data in QLC mode.
>> *Ok, true.... yes....*
>> A 100 GB SLC cache would reduce the capacity of a 1 TB SSD to 700 GB (600 GB
>> in 150 tn QLC cells plus 100 GB in 100 tn SLC cells).
>> *Ahh! you mean that SLC capacity for speeding up the QLC disks, is obtained
>> from each single layer of the QLC?.*
There are no specific SLC cells. A fraction of the QLC capable cells is only
written with only 1 instead of 4 bits. This is a much simpler process, since
there are only 2 charge levels per cell that are used, while QLC uses 16 charge
levels, and you can only add charge (must not overshoot), therefore only small
increments are added until the correct value can be read out).

But since SLC cells take away specified capacity (which is calculated assuming
all cells hold 4 bits each, not only 1 bit), their number is limited and
shrinks as demand for QLC cells grows.

The advantage of the SLC cache is fast writes, but also that data in it may
have become stale (trimmed) and thus will never be copied over into a QLC
block. But as the SSD fills and the size of the SLC cache shrinks, this
capability will be mostly lost, and lots of very short lived data is stored in
QLC cells, which will quickly become partially stale and thus needing
compaction as explained above.

>> Therefore, the fraction of the cells used as an SLC cache is reduced when it
>> gets full (e.g. ~1 TB in ~250 tn QLC cells, plus 6 GB in 6 tn SLC cells).
>> *Sorry I don't get this last sentence... don't understand it because I don't
>> really know the meaning of tn... *
>> *but I think I'm getting the idea if you say that each QLC layer, has it's
>> own SLC cache obtained from the disk space avaiable for each QLC layer....*
>> And with less SLC cells available for short term storage of data the
>> probability of data being copied to QLC cells before the irrelevant messages
>> have been deleted is significantly increased. And that will again lead to
>> many more blocks with "holes" (deleted messages) in them, which then need to
>> be copied possibly multiple times to compact them.
>> *If I correct above, I think I got the idea yes....*
>>>     (on many SSDs, I have no numbers for this
>>>     particular device), and thus the amount of data that can be
>>>     written at single cell speed shrinks as the SSD gets full.
>>>     I have just looked up the size of the SLC cache, it is specified
>>>     to be 78 GB for the empty SSD, 6 GB when it is full (for the 2 TB
>>>     version, smaller models will have a smaller SLC cache).
>>>     *Assuming you were talking about the cache for compensating speed we
>>>     previously commented, I should say these are the 870 QVO but the 8TB
>>>     version. So they should have the biggest cache for compensating the
>>>     speed issues...*
>> I have looked up the data: the larger versions of the 870 QVO have the same
>> SLC cache configuration as the 2 TB model, 6 GB minimum and up to 72 GB more
>> if there are enough free blocks.
>> *Ours one is the 8TB model so I assume it could have bigger limits. The
>> disks are mostly empty, really.... so... for instance....*
>> *zpool list*
>> *root_dataset  448G  2.29G   446G        -         -     1%     0%  1.00x 
>> ONLINE  -*
>> *mail_dataset  58.2T  11.8T  46.4T        -         -    26%    20%  1.00x 
>> ONLINE  -*
Ok, seems you have got 10 * 8 TB in a raidz2 configuration.

Only 20% of the mail dataset is in use, the situation will become much worse
when the pool will fill up!

>> *I suppose fragmentation affects too....*
On magnetic media fragmentation means that a file is spread out over the disk
in a non-optimal way, causing access latencies due to seeks and rotational
delay. That kind of fragmentation is not really relevant for SSDs, which allow
for fast random access to the cells.

And the FRAG value shown by the "zpool list" command is not about fragmentation
of files at all, it is about the structure of free space. Anyway less relevant
for SSDs than for classic hard disk drives.

>>>     But after writing those few GB at a speed of some 500 MB/s (i.e.
>>>     after 12 to 150 seconds), the drive will need several minutes to
>>>     transfer those writes to the quad-level cells, and will operate
>>>     at a fraction of the nominal performance during that time.
>>>     (QLC writes max out at 80 MB/s for the 1 TB model, 160 MB/s for the
>>>     2 TB model.)
>>>     *Well we are in the 8TB model. I think I have understood what you wrote
>>>     in previous paragraph. You said they can be fast but not constantly,
>>>     because later they have to write all that to their perpetual storage
>>>     from the cache. And that's slow. Am I wrong?. Even in the 8TB model you
>>>     think Stefan?.*
>> The controller in the SSD supports a given number of channels (e.g 4), each
>> of which can access a Flash chip independently of the others. Small SSDs
>> often have less Flash chips than there are channels (and thus a lower
>> throughput, especially for writes), but the larger models often have more
>> chips than channels and thus the performance is capped.
>> *This is totally logical. If a QVO disk would outperform best or similar
>> than an Intel without consequences.... who was going to buy a expensive
>> Intel enterprise?.*
The QVO is bandwidth limited due to the SATA data rate of 6 Mbit/s anyway, and
it is optimized for reads (which are not significantly slower than offered by
the TLC models). This is a viable concept for a consumer PC, but not for a server.
>> In the case of the 870 QVO, the controller supports 8 channels, which allows
>> it to write 160 MB/s into the QLC cells. The 1 TB model apparently has only
>> 4 Flash chips and is thus limited to 80 MB/s in that situation, while the
>> larger versions have 8, 16, or 32 chips. But due to the limited number of
>> channels, the write rate is limited to 160 MB/s even for the 8 TB model.
>> *Totally logical Stefan...*
>> If you had 4 * 2 TB instead, the throughput would be 4 * 160 MB/s in this limit.
>>>     *The main problem we are facing is that in some peak moments, when the
>>>     machine serves connections for all the instances it has, and only as
>>>     said in some peak moments... like the 09am or the 11am.... it seems the
>>>     machine becomes slower... and like if the disks weren't able to serve
>>>     all they have to serve.... In these moments, no big files are moved...
>>>     but as we have 1800-2000 concurrent imap connections... normally they
>>>     are doing each one... little changes in their mailbox. Do you think
>>>     perhaps this disks then are not appropriate for this kind of usage?-*
>> I'd guess that the drives get into a state in which they have to recycle
>> lots of partially free blocks (i.e. perform kind of a garbage collection)
>> and then three kinds of operations are competing with each other:
>>  1. reads (generally prioritized)
>>  2. writes (filling the SLC cache up to its maximum size)
>>  3. compactions of partially filled blocks (required to make free blocks
>>     available for re-use)
>> Writes can only proceed if there are sufficient free blocks, which on a
>> filled SSD with partially filled erase blocks means that operations of type
>> 3. need to be performed with priority to not stall all writes.
>> My assumption is that this is what you are observing under peak load.
>> *It could be although the disks are not filled.... the pool are at 20 or 30%
>> of capacity and fragmentation from 20%-30% (as zpool list states).*
Yes, and that means that your issues will become much more critical over time
when the free space shrinks and garbage collections will be required at an even
faster rate, with the SLC cache becoming less and less effective to weed out
short lived files as an additional factor that will increase write amplification.
>>>     And cheap SSDs often have no RAM cache (not checked, but I'd be
>>>     surprised if the QVO had one) and thus cannot keep bookkeeping date
>>>     in such a cache, further limiting the performance under load.
>>>     *This brochure
>>>     (
>>>     and the datasheet
>>>     sais if I have read properly, the 8TB drive has 8GB of ram?. I assume
>>>     that is what they call the turbo write cache?.*
>> No, the turbo write cache consists of the cells used in SLC mode (which can
>> be any cells, not only cells in a specific area of the flash chip).
>> *I see I see....*
>> The RAM is needed for fast lookup of the position of data for reads and of
>> free blocks for writes.
>> *Our ones... seem to have 8GB LPDDR4 of ram.... as datasheet states....*
Yes, and it makes sense that the RAM size is proportional to the capacity since
a few bytes are required per addressable data block.

If the block size was 8 KB the RAM could hold 8 bytes (e.g. a pointer and some
status flags) for each logically addressable block. But there is no information
about the actual internal structure of the QVO that I know of.

>> *I see.... It's extremely misleading you know... because... you can copy
>> five mailboxes of 50GB concurrently for instance.... and you flood a gigabit
>> interface copying (obviously because disks can keep that throughput)... but
>> later.... you see... you are in an hour that yesterday, and even 4 days
>> before you have not had any issues... and that day... you see the commented
>> issue... even not being exactly at a peak hour (perhaps is two hours later
>> the peak hour even)... or... but I wasn't noticing about all things you say
>> in this email....*
>> I have seen advice to not use compression in a high load scenario in some
>> other reply.
>> I tend to disagree: Since you seem to be limited when the SLC cache is
>> exhausted, you should get better performance if you compress your data. I
>> have found that zstd-2 works well for me (giving a significant overall
>> reduction of size at reasonable additional CPU load). Since ZFS allows to
>> switch compressions algorithms at any time, you can experiment with
>> different algorithms and levels.
>> *I see... you say compression should be enabled.... The main reason because
>> we have not enabled it yet, is for keeping the system the most near possible
>> to config defaults... you know... for later being able to ask in this
>> mailing lists if we have an issue... because you know... it would be far
>> more easier to ask about something strange you are seeing when that strange
>> thing is near to a well tested config, like the config by default....*
>> *But now you say Stefan... if you switch between compression algorithms you
>> will end up with a mix of different files compressed in a different
>> manner... that is not a bit disaster later?. Doesn't affect performance in
>> some manner?.*
The compression used is stored in the per file information, each file in a
dataset could have been written with a different compression method and level.
Blocks are independently compressed - a file level compression may be more
effective. Large mail files will contain incompressible attachments (already
compressed), but in base64 encoding. This should allow a compression ratio of
~1,3. Small files will be plain text or HTML, offering much better compression
>> One advantage of ZFS compression is that it applies to the ARC, too. And a
>> compression factor of 2 should easily be achieved when storing mail (not for
>> .docx, .pdf, .jpg files though). Having more data in the ARC will reduce the
>> read pressure on the SSDs and will give them more cycles for garbage
>> collections (which are performed in the background and required to always
>> have a sufficient reserve of free flash blocks for writes).
>> *We would use I assume the lz4... which is the less "expensive" compression
>> algorithm for the CPU... and I assume too for avoiding delay accessing
>> data... do you recommend another one?. Do you always recommend compression
>> then?.*
I'd prefer zstd over lz4 since it offers a much higher compression ratio.

Zstd offers higher compression ratios than lz4 at similar or better
decompression speed, but may be somewhat slower compressing the data. But in my
opinion this is outweighed by the higher effective amount of data in the
ARC/L2ARC possible with zstd.

For some benchmarks of different compression algorithms available for ZFS and
compared to uncompressed mode see the extensive results published by Jude Allan:

The SQL benchmarks might best resemble your use case - but remember that a significant reduction of the amount of data being written to the SSDs might be more important than the highest transaction rate, since your SSDs put a low upper limit on that when highly loaded.

>> I'd give it a try - and if it reduces your storage requirements by 10% only,
>> then keep 10% of each SSD unused (not assigned to any partition). That will
>> greatly improve the resilience of your SSDs, reduce the write-amplification,
>> will allow the SLC cache to stay at its large value, and may make a large
>> difference to the effective performance under high load.
>> *But when you enable compression... only gets compressed the new data
>> modified or entered. Am I wrong?.*
Compression is per file system data block (at most 1 MB if you set the
blocksize to that value). Each such block is compressed independently of all
others, to not require more than 1 block to be read and decompressed when
randomly reading a file. If a block does not shrink when compressed (it may
contain compressed file data) the block is written to disk as-is (uncompressed).
>> **
>> *By the way, we have more or less 1/4 of each disk used (12 TB allocated in
>> a poll stated by zpool list, divided between 8 disks of 8TB...)... do you
>> think we could be suffering on write amplification and so... having a so
>> little disk space used in each disk?.*
Your use case will cause a lot of garbage collections and this particular high
write amplification values.
>> Regards, STefan
>> *Hey mate, your mail is incredible. It has helped as a lot. Can we invite
>> you a cup of coffee or a beer through Paypal or similar?. Can I help you in
>> some manner?.*
Thanks, I'm glad to help, and I'd appreciate to hear whether you get your setup
optimized for the purpose (and how well it holds up when you approach the
capacity limits of your drives).

I'm always interested in experience of users with different use cases than I
have (just being a developer with too much archived mail and media collected
over a few decades).

Regards, STefan