oh I don't know if it helps, but the file types and content between the old system and the new one are the exact same type (file type/structure did not changed between the two)
Last edited:
-
Important Announcement for the TrueNAS Community.
The TrueNAS Community has now been moved. This forum has become READ-ONLY for historical purposes. Please feel free to join us on the new TrueNAS Community Forums
Can't get why scrub is so slow / replication failing during slow scrub
- Thread starter Ulysse_31
- Start date
For the zpool iostat question ... well it seems to reflect the scrubbing :
Code:
zpool iostat STORAGE1 1 1000
capacity operations bandwidth
pool alloc free read write read write
---------- ----- ----- ----- ----- ----- -----
STORAGE1 57.6T 70.6T 2.08K 181 47.1M 5.98M
STORAGE1 57.6T 70.6T 5.82K 0 33.9M 0
STORAGE1 57.6T 70.6T 4.76K 0 66.1M 0
STORAGE1 57.6T 70.6T 4.95K 0 88.6M 0
STORAGE1 57.6T 70.6T 52 1.09K 211K 14.4M
STORAGE1 57.6T 70.6T 3.77K 23 24.2M 166K
STORAGE1 57.6T 70.6T 5.24K 0 33.5M 0
STORAGE1 57.6T 70.6T 6.69K 0 45.2M 0
STORAGE1 57.6T 70.6T 6.58K 0 42.6M 0
STORAGE1 57.6T 70.6T 6.54K 0 43.5M 0
STORAGE1 57.6T 70.6T 6.41K 0 47.1M 0
STORAGE1 57.6T 70.6T 6.75K 0 46.0M 0
STORAGE1 57.6T 70.6T 7.09K 0 44.5M 0
STORAGE1 57.6T 70.6T 6.91K 0 38.7M 0
STORAGE1 57.6T 70.6T 6.85K 0 37.3M 0
STORAGE1 57.6T 70.6T 6.75K 0 38.9M 0
STORAGE1 57.6T 70.6T 5.27K 0 30.0M 0
STORAGE1 57.6T 70.6T 5.10K 0 105M 0
STORAGE1 57.6T 70.6T 2.66K 134 44.3M 1.74M
STORAGE1 57.6T 70.6T 2.56K 300 16.7M 1.77M
STORAGE1 57.6T 70.6T 4.69K 0 29.3M 0
STORAGE1 57.6T 70.6T 6.76K 0 46.0M 0
STORAGE1 57.6T 70.6T 6.62K 0 41.8M 0
STORAGE1 57.6T 70.6T 6.72K 0 47.3M 0
STORAGE1 57.6T 70.6T 6.49K 0 49.0M 0
STORAGE1 57.6T 70.6T 6.84K 0 48.7M 0
STORAGE1 57.6T 70.6T 7.34K 0 46.1M 0
STORAGE1 57.6T 70.6T 7.11K 0 40.4M 0
STORAGE1 57.6T 70.6T 6.86K 0 37.3M 0
STORAGE1 57.6T 70.6T 6.74K 0 38.8M 0
STORAGE1 57.6T 70.6T 5.44K 0 32.2M 0
STORAGE1 57.6T 70.6T 4.21K 0 90.7M 0
STORAGE1 57.6T 70.6T 1.22K 282 25.1M 2.86M
STORAGE1 57.6T 70.6T 2.36K 554 15.4M 10.2M
STORAGE1 57.6T 70.6T 4.50K 0 27.7M 0
STORAGE1 57.6T 70.6T 6.69K 0 45.5M 0
STORAGE1 57.6T 70.6T 6.65K 0 41.8M 0
STORAGE1 57.6T 70.6T 6.76K 0 47.8M 0
STORAGE1 57.6T 70.6T 6.42K 0 45.6M 0
STORAGE1 57.6T 70.6T 6.42K 0 45.8M 0
STORAGE1 57.6T 70.6T 7.32K 0 46.4M 0
STORAGE1 57.6T 70.6T 7.19K 0 41.7M 0
STORAGE1 57.6T 70.6T 6.86K 0 36.8M 0
STORAGE1 57.6T 70.6T 6.89K 0 39.8M 0
STORAGE1 57.6T 70.6T 5.45K 0 32.3M 0
STORAGE1 57.6T 70.6T 5.06K 0 104M 0
STORAGE1 57.6T 70.6T 3.51K 0 49.9M 0
STORAGE1 57.6T 70.6T 798 703 5.19M 9.70M
STORAGE1 57.6T 70.6T 4.06K 0 25.5M 0
STORAGE1 57.6T 70.6T 5.95K 0 39.2M 0
STORAGE1 57.6T 70.6T 6.71K 0 44.2M 0
Please note that the pool does nothing particular during the day ... nothing at all in fact ... it only "works" during night while replicating ...
Last edited by a moderator:
- Joined
- Feb 15, 2014
- Messages
- 20,194
Can you get us the output with -q?
Sure, here it is :
Code:
zpool iostat -q STORAGE1 1 1000
capacity operations bandwidth syncq_read syncq_write asyncq_read asyncq_write scrubq_read trimq_write
pool alloc free read write read write pend activ pend activ pend activ pend activ pend activ pend activ
---------- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- -----
STORAGE1 57.7T 70.6T 2.10K 180 47.0M 5.95M 0 0 0 0 0 0 0 0 5.54K 34 0 0
STORAGE1 57.7T 70.6T 6K 0 34.6M 0 0 0 0 0 0 0 0 0 5.41K 36 0 0
STORAGE1 57.7T 70.6T 5.12K 0 47.8M 0 0 0 0 0 0 0 0 0 5.56K 30 0 0
STORAGE1 57.7T 70.6T 4.86K 0 117M 0 0 0 0 0 0 0 0 0 869 20 0 0
STORAGE1 57.7T 70.6T 793 522 6.80M 7.51M 0 0 0 0 0 0 24 9 0 0 0 0
STORAGE1 57.7T 70.6T 6.80K 23 54.1M 165K 0 0 0 0 0 0 0 0 8.09K 36 0 0
STORAGE1 57.7T 70.6T 6.92K 0 55.2M 0 0 0 0 0 0 0 0 0 7.19K 36 0 0
STORAGE1 57.7T 70.6T 7.27K 0 47.2M 0 0 0 0 0 0 0 0 0 6.21K 36 0 0
STORAGE1 57.7T 70.6T 7.02K 0 42.3M 0 0 0 0 0 0 0 0 0 6.55K 36 0 0
STORAGE1 57.7T 70.6T 7.02K 0 39.5M 0 0 0 0 0 0 0 0 0 6.67K 36 0 0
STORAGE1 57.7T 70.6T 6.75K 0 36.9M 0 0 0 0 0 0 0 0 0 5.74K 35 0 0
STORAGE1 57.7T 70.6T 6.38K 0 35.2M 0 0 0 0 0 0 0 0 0 4.86K 36 0 0
STORAGE1 57.7T 70.6T 5.05K 0 30.2M 0 0 0 0 0 0 0 0 0 6.25K 35 0 0
STORAGE1 57.7T 70.6T 4.00K 0 94.5M 0 0 0 0 0 0 0 0 0 1.93K 9 0 0
STORAGE1 57.7T 70.6T 895 367 21.0M 3.71M 0 0 0 0 0 0 0 1 0 0 0 0
STORAGE1 57.7T 70.6T 5.00K 469 40.5M 10.6M 0 0 0 0 0 0 0 0 7.96K 36 0 0
STORAGE1 57.7T 70.6T 6.98K 0 54.8M 0 0 0 0 0 0 0 0 0 7.04K 36 0 0
STORAGE1 57.7T 70.6T 7.18K 0 50.3M 0 0 0 0 0 0 0 0 0 7.10K 35 0 0
STORAGE1 57.7T 70.6T 7.20K 0 42.9M 0 0 0 0 0 0 0 0 0 6.83K 36 0 0
STORAGE1 57.7T 70.6T 7.15K 0 41.1M 0 0 0 0 0 0 0 0 0 5.90K 36 0 0
STORAGE1 57.7T 70.6T 6.86K 0 37.3M 0 0 0 0 0 0 0 0 0 6.06K 33 0 0
STORAGE1 57.7T 70.6T 6.51K 0 36.5M 0 0 0 0 0 0 0 0 0 5.17K 36 0 0
STORAGE1 57.7T 70.6T 5.17K 0 29.4M 0 0 0 0 0 0 0 0 0 6.45K 36 0 0
STORAGE1 57.7T 70.6T 5.21K 0 110M 0 0 0 0 0 0 0 0 0 2.93K 15 0 0
STORAGE1 57.7T 70.6T 2.57K 100 37.5M 1.35M 0 0 0 0 0 0 9 2 0 0 0 0
STORAGE1 57.7T 70.6T 4.15K 363 33.7M 3.76M 0 0 0 0 0 0 0 0 8.14K 36 0 0
STORAGE1 57.7T 70.6T 7.00K 0 55.1M 0 0 0 0 0 0 0 0 0 7.20K 36 0 0
STORAGE1 57.7T 70.6T 7.10K 0 51.8M 0 0 0 0 0 0 0 0 0 6.91K 36 0 0
STORAGE1 57.7T 70.6T 7.29K 0 43.3M 0 0 0 0 0 0 0 0 0 5.55K 36 0 0
STORAGE1 57.7T 70.6T 7.02K 0 41.2M 0 0 0 0 0 0 0 0 0 6.98K 36 0 0
STORAGE1 57.7T 70.6T 6.93K 0 37.6M 0 0 0 0 0 0 0 0 0 6.29K 36 0 0
STORAGE1 57.7T 70.6T 6.50K 0 35.7M 0 0 0 0 0 0 0 0 0 5.42K 34 0 0
STORAGE1 57.7T 70.6T 5.47K 0 31.8M 0 0 0 0 0 0 0 0 0 5.11K 36 0 0
STORAGE1 57.7T 70.6T 4.96K 0 88.4M 0 0 0 0 0 0 0 0 0 2.86K 24 0 0
STORAGE1 57.7T 70.6T 4.42K 0 69.7M 0 0 0 0 0 0 0 0 0 151 9 0 0
STORAGE1 57.7T 70.6T 1.32K 443 8.51M 5.33M 0 0 0 0 0 0 0 0 5.09K 36 0 0
STORAGE1 57.7T 70.6T 7.11K 0 57.4M 0 0 0 0 0 0 0 0 0 7.88K 36 0 0
STORAGE1 57.7T 70.6T 6.79K 0 54.4M 0 0 0 0 0 0 0 0 0 7.40K 36 0 0
STORAGE1 57.7T 70.6T 7.27K 0 45.2M 0 0 0 0 0 0 0 0 0 7.35K 36 0 0
STORAGE1 57.7T 70.6T 6.93K 0 41.8M 0 0 0 0 0 0 0 0 0 7.17K 36 0 0
STORAGE1 57.7T 70.6T 6.97K 0 38.9M 0 0 0 0 0 0 0 0 0 6.64K 36 0 0
STORAGE1 57.7T 70.6T 6.71K 0 36.5M 0 0 0 0 0 0 0 0 0 5.83K 29 0 0
STORAGE1 57.7T 70.6T 6.22K 0 34.5M 0 0 0 0 0 0 0 0 0 5.25K 36 0 0
STORAGE1 57.7T 70.6T 5.04K 0 30.3M 0 0 0 0 0 0 0 0 0 5.76K 33 0 0
STORAGE1 57.7T 70.6T 3.81K 0 94.0M 0 0 0 0 0 0 0 0 0 1.64K 9 0 0
STORAGE1 57.7T 70.6T 801 358 19.3M 5.68M 0 0 0 0 0 0 65 11 0 0 0 0
STORAGE1 57.7T 70.6T 5.96K 170 47.6M 1.47M 0 0 0 0 0 0 0 0 7.64K 36 0 0
STORAGE1 57.7T 70.6T 6.99K 0 54.7M 0 0 0 0 0 0 0 0 0 7.07K 36 0 0
STORAGE1 57.7T 70.6T 7.14K 0 48.8M 0 0 0 0 0 0 0 0 0 6.61K 36 0 0
STORAGE1 57.7T 70.6T 7.08K 0 42.6M 0 0 0 0 0 0 0 0 0 5.88K 36 0 0
STORAGE1 57.7T 70.6T 7.00K 0 40.0M 0 0 0 0 0 0 0 0 0 6.26K 36 0 0
Don't know how many (how much time) you would like ... sorry for the spam if its too big ^^' and tell me if you need more (maybe as an attachment it would be better).
Last edited by a moderator:
WI_Hedgehog
Guru
- Joined
- Jun 15, 2022
- Messages
- 674
Nailed it.
IOPS (Input/Output Operations Per Second) Definition
The definition of IOPS defined and explained in simple language.
Hello ^^Nailed it.
IOPS (Input/Output Operations Per Second) Definition
The definition of IOPS defined and explained in simple language.techterms.com
I have no issue with those definitions, I think that the main issue here is the usage of some words, and some shortcuts that some may think when earing/reading something ...
I'll try to reformulate, again, just to be sure that no misunderstanding will persist (again) ^^ :
A given pool, by its setup / config, will have a given number of IOPS (Input and Output Per Seconds) on a given circumstance. Those IOPS can be quantified with a value which are inputs and outputs usable per seconds.
when operations (reading/writing) will be made on that pool, that amount, the inputs and outputs per seconds done on the pool, will be spread accross the drives that are inside the pool ... I could also say : during one second, inputs and outputs will happen over the drives composing that pool ...
That is why ... IOPS of a pool are spread accross the drives of that pool ... and that was exactly what I meant : not more, not less
I DID NOT meant that IOPS would be in some way increasing or cumulated or ... any other thing
Hope that helps removing some misunderstanding that some may had reading my early posts ^^' (really hope ... I'm honestly getting tired on this subject).
Last edited:
- Joined
- Feb 15, 2014
- Messages
- 20,194
The length is fine, but I would appreciate it if you could use CODE Tags instead of QUOTE next time, to preserve whitespace. I've edited your posts accordingly.
Now, as for interpreting the output... You're 100% IOPS-bound, but that's not a surprise. I'm not sure if there's an easy way to explicitly get figures for actual IOPS, rather than sampling the queue lengths. I guess we can examine it at the disk level with
If the disks show 100-200 IOPS each, that's a sign that the problem is that the pool is too fragmented (keep in mind that the fragmentation value reported by ZFS is only Free Space Fragmentation, which impacts write performance but does not really impact reads directly).
If the disks show 10-20 IOPS each, that's a sign that the problem is more on the hardware layer - with the prime suspect at this time being the HBA.
As a point of comparison (sadly without direct iostat comparisons, because the machine is running a Pre-OpenZFS merger version), a pool with a single 12-wide RAIDZ2, with mostly large blocks - though with some non-performance-critical zvols thrown in, plus a lot of tiny-to-smallish files in the mix (maybe like 10 small files for each huge one) - with 10 TB HDDs and close to 80% full, is right now pushing well over 1.5 GB/s of issued scrub reads - and it's using up IOPS to fill up the scan buffer so that it can issue the reads sequentially. This corresponds to high 50s/low 60s read operations per second on every disk, plus a handful of write ops.
Now, as for interpreting the output... You're 100% IOPS-bound, but that's not a surprise. I'm not sure if there's an easy way to explicitly get figures for actual IOPS, rather than sampling the queue lengths. I guess we can examine it at the disk level with
iostat -x.If the disks show 100-200 IOPS each, that's a sign that the problem is that the pool is too fragmented (keep in mind that the fragmentation value reported by ZFS is only Free Space Fragmentation, which impacts write performance but does not really impact reads directly).
If the disks show 10-20 IOPS each, that's a sign that the problem is more on the hardware layer - with the prime suspect at this time being the HBA.
As a point of comparison (sadly without direct iostat comparisons, because the machine is running a Pre-OpenZFS merger version), a pool with a single 12-wide RAIDZ2, with mostly large blocks - though with some non-performance-critical zvols thrown in, plus a lot of tiny-to-smallish files in the mix (maybe like 10 small files for each huge one) - with 10 TB HDDs and close to 80% full, is right now pushing well over 1.5 GB/s of issued scrub reads - and it's using up IOPS to fill up the scan buffer so that it can issue the reads sequentially. This corresponds to high 50s/low 60s read operations per second on every disk, plus a handful of write ops.
They're not spread. One request from ZFS's higher layers will cause one request per disk, with all disks contributing to the same, single request. Yes, they're smaller, but that does not help you, since the disks are limited by how many requests they can serve, regardless of how large they are.
So a quick update about the slow scrubbing running actually on our TrueNAS node with issues :
Speed of scrub is slowly decreasing (we started at around 50MBytes per sec ... we are now at 38.8 ...) ... no errors for now ...
The estimated time increases also ... seems to turn into a never ending story ...
For now, I have no other hypothesis/clues than the firmware card (maybe ... but will wait that the scrub ends to plan an export/import of the pool ... since it seems we should not be accessing disk data during the flashing ...)
I really cannot understand why those poor performances ... we have a non-TrueNAS with the exact same data type and with bigger data usage with more than 5 times the performances of this ...
Speed of scrub is slowly decreasing (we started at around 50MBytes per sec ... we are now at 38.8 ...) ... no errors for now ...
The estimated time increases also ... seems to turn into a never ending story ...
For now, I have no other hypothesis/clues than the firmware card (maybe ... but will wait that the scrub ends to plan an export/import of the pool ... since it seems we should not be accessing disk data during the flashing ...)
I really cannot understand why those poor performances ... we have a non-TrueNAS with the exact same data type and with bigger data usage with more than 5 times the performances of this ...
OMG ... do you mean only one IO is made per sec on a pool ? or do you mean in one sec only one drive will be asked on a pool ?
...
doesn't those two nonsense question gives you an hint on what I wrote ?
Last edited:
roger that will do ^^
strange since zpool status -v shows no frag ... AS you said => it shows only free space fragmentation ... ok but how can a drive be fragmented on a COW system that does not delete data ?
UPDATE: sorry, read a bit too quick, apologize on that one ... ok you meant a comparison ... again ... we are on a 44% used pool
Last edited:
hmmm ... not sure but man zpool-iostat under TrueNAS does not mention about a -x option ... ?
Oh ... you meant iostat directly this time ^^ let me look at it ^^
- Joined
- Feb 15, 2014
- Messages
- 20,194
No, not at all. But one operation on the vdev means that every disk needs to do the same operation. Hence why you have the same IOPS capability as a single disk.
You're not wrong, in the sense that your case does not sound like it should be pathological. It's not completely out of the question, but it doesn't really fit the picture you've painted.
But I never meant that a pool would have more, or less (or even anything related) IOPS capability than one drive ! that IS my point !
I just said that the inputs and outputs per seconds happening on a pool during a system operation (like a scrub), are spread among the 12 drives that composes the pool. I never, ever talk about the quantity of IOPS.
I said that the "n" IOs happening per sec (referring notably during the scrub operation) are spread between the 12 drives on the pool ...
Really sorry if you cannot understand what I'm trying to explain ... did my best ... ^^
Could you be more explicit on this one ? what does not fit what ?
- Joined
- Feb 15, 2014
- Messages
- 20,194
Your description does not suggest a workload that would explain the poor performance.
- Joined
- Jul 12, 2022
- Messages
- 3,222
@Ulysse_31 I have seen the first page of this thread and I suggest reading the following resource.
www.truenas.com
ZFS Storage Pool Layout
This resource was originally created by user: @Davvo on the TrueNAS Community Forums Archive. https://www.truenas.com/community/resources/zfs-storage-pool-layout.201/download [1] This amazing document, created by iXsystems in February 2022 as a “White Paper”, cleanly explains how to qualify...
www.truenas.com
Hi Davvo ^^@Ulysse_31 I have seen the first page of this thread and I suggest reading the following resource.
ZFS Storage Pool Layout
This resource was originally created by user: @Davvo on the TrueNAS Community Forums Archive. https://www.truenas.com/community/resources/zfs-storage-pool-layout.201/download [1] This amazing document, created by iXsystems in February 2022 as a “White Paper”, cleanly explains how to qualify...
Since now my account is not a "rookie" account anymore ^^' I'll be able to update that first post and add some additional informations that were discussed afterwards in an UPDATE status ^^
But let me resume it here, I'll use this occasion to give even more informations:
This server is a "droppin replacement" of an already existing zfs "backup" replication server: its config / shape / potential bottlenecks ... following our type of data : we setup the first "basic shape & config" of this server role in ... 2014 ... by "basic shape" I'm talking about the choice of using a poweredge server with a powervault SAS disk bay in raidz2, at that time, the OS was a freebsd 9, and our first tool scripts to make zfs syncs.
In 2017, since the production node was a solaris 11, and to avoid getting too distant versions of zfs, we decided to move the "zfs replication server" also to a solaris 11 OS.
So we built again, same structure, same shape : a poweredge server, a powervault drive bay, this time 12x 4Tb SAS drives, since this version worked just fine in tandem with the production server, we decided to prolong the lifetime of this last one a first time: when it arrived at 90% pool usage in 2019, we added a new powervault drive bay in daisy chain, and added a new raidz2 vdev to the pool. And we double down in 2021 with a third bay, and a third vdev, again, when it arrived at 90% pool usage.
With both hosts, either the freebsd or the solaris, we never had bottleneck issues during scrubs, those two servers did there job very well.
We recently moved the production server to a TrueNAS server, a M40-HA, in which we replicated the same data structure / data usage that was in the old production server also ... so the type of data is still the exact same type.
That is why we decided to give a try on a "zfs replication server" based on TrueNAS. So again, we took the same setup & hardware profile : a poweredge server, a powervault SAS drive bay, and 12x 12Tb drives this time.
Right now, on its last iteration, the Solaris 11 version, which is right now filled with 6 years of data retention, and 111Tb of data (85% pool usage) we are making scrubs with speeds of 489MBytes per sec.
I can totally understand that "depending on the data type and desired IO load, we need to select the pool & hardware setup accordingly", that of course makes total sense.
But we are talking here of an hardware setup and profile that was tested for our usage since pretty much time and we NEVER had issues ^^" so it is well sized for our usage.
I would like to add an extra side note : We are using zfs since ... a while now ... I touched my first zfs filesystem in solaris 8 (was in ... 2008? ^^' ) while it was owned by Sun ... and here in the company I work for since 2014. We use zfs in various other contexts, we discovered and had goods and bads (the good of snapshoting and replicating ... the bad of dedup :p ... stuff like that ...) We use zfs under linux with ZoL ... we use it in a proxmox cluster environment ... we even use zfs in openindiana on an archival system ... and ... on all those years of usage ... AND within our specific usage we never had a bottleneck like that during scrub ...
But let me add an update text into the first post to avoid other misleading conclusions ^^'
Last edited:
Nor FreeBSD 9 right ^^' but I would not expect an 6 years old version of zfs (v31 /v28 freebsd / 5000 openzfs) to be more efficient on our "basic" usage ^^' than actual 2023 openzfs version ...
and this is also why I'm clearly more incline on searching either a firmware (card or drive) potential issue, a bad cable or somewhat any hardware related issue ...
But my main concern in this is the difficulty finding real error message => for now, unless those messages from the SAS card, that are not an error but it being looping may be suspicious ... I do not have anything for now ...
if it was errors happing at read or write ... I would expect to have timeout errors somewhere ... oh well ... again, I'll have to wait for the scrub ending to plan a firmware card upgrade ...
