Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
cMP Slow AFPS 4K Write performance
- Thread starter handheldgames
- WikiPost WikiPost
- Start date
- Sort by reaction score
- Status
If you want to enable [4K block] I/O on controller level. it seems the case. But make sure the firmware of your SAS controller allows you to do so. Not all SAS controller have such option. I don't think it is available on those cheap models. Try ATTO or Acrea (enterprise grade SAS controllers).
Reading this thread (and a few others before it) almost made me cry. I got an iMac Pro for 4-5 months now and lately figured I'll just put its advertised speeds to the test. Imagine my surprise...
I also downloaded the new-ish Sensei app (website is literally sensei.app) and enabled TRIM from there, waited 5 minutes for kernel modules cache to rebuild and rebooted. Re-ran benchmarks, no appreciable difference except that both 4K write speeds went to 14 MB/s.
I have NOT rebooted in single-user mode and have NOT ran `fsck -fy` though. I hear it might help due to actually flagging a lot of areas as properly freed, or something.
As a programmer database performance on my workstation is very important. I haven't seen any problems btw; but the 2x difference between actual sequential write speeds (1.7GB/s) vs the marketed one (3.3GB/s) is worrying. Anyone has an idea where is the discrepancy coming from? I got the 10-core variant with 64GB RAM and (obviously, as seen on the screenshots) the 2TB SSD.
I keep seeing normal user-grade SSDs outperforming my supposed beastly SSD in the 4K write tests. While I haven't noticed any actual slowdown in heavy workloads, I am still kind of worried.
(Maybe I should run the `fio` tests, still haven't done that.)
I also downloaded the new-ish Sensei app (website is literally sensei.app) and enabled TRIM from there, waited 5 minutes for kernel modules cache to rebuild and rebooted. Re-ran benchmarks, no appreciable difference except that both 4K write speeds went to 14 MB/s.
I have NOT rebooted in single-user mode and have NOT ran `fsck -fy` though. I hear it might help due to actually flagging a lot of areas as properly freed, or something.
As a programmer database performance on my workstation is very important. I haven't seen any problems btw; but the 2x difference between actual sequential write speeds (1.7GB/s) vs the marketed one (3.3GB/s) is worrying. Anyone has an idea where is the discrepancy coming from? I got the 10-core variant with 64GB RAM and (obviously, as seen on the screenshots) the 2TB SSD.
I keep seeing normal user-grade SSDs outperforming my supposed beastly SSD in the 4K write tests. While I haven't noticed any actual slowdown in heavy workloads, I am still kind of worried.
(Maybe I should run the `fio` tests, still haven't done that.)
Last edited:
Figured I'll run Blackmagic as well:
Both reads and writes seem about 350MB/s below advertised speeds (2.45GB/s vs 2.8GB/s read -- and 2.95GB/s vs 3.3GB/s write) but this looks a bit better.
I suppose that the benchmarking tools in general haven't been kept up to speed with such hardware plus the APFS filesystem peculiarities. Or maybe Blackmagic tests unrealistic workflows (seems to only test sequential read/write) to make the owner of the machine feel better.
EDIT: This early review seems to confirm my numbers perfectly.
Both reads and writes seem about 350MB/s below advertised speeds (2.45GB/s vs 2.8GB/s read -- and 2.95GB/s vs 3.3GB/s write) but this looks a bit better.
I suppose that the benchmarking tools in general haven't been kept up to speed with such hardware plus the APFS filesystem peculiarities. Or maybe Blackmagic tests unrealistic workflows (seems to only test sequential read/write) to make the owner of the machine feel better.
EDIT: This early review seems to confirm my numbers perfectly.
Last edited:
The "something" is that if an SSD disk is not enabled for TRIM, file deletions are not passed through to the disk, so that it does not (and can not) know that those file allocations are now "free" space.
OSX might tell you that the disk is 50% full, but the garbage collector might see it as 90% full. This makes the garbage collector do extra work, which slows writes and causes unnecessary wear on the disk.
Enabling TRIM says that future file deletes will be properly marked as free space - but it does not actually free up any old deleted areas.
'fsck -fy' gets the OS and the disk to agree on how much (and which) space is free.
Last edited:
Sure. I am just kind of afraid if I will be able to even reboot afterwards. I guess I am chickening out about it at the moment.
That's why we always keep complete and up-to-date backups, right?
I've read many horror stories about how useless TimeMachine turned out to be when crap really did hit the fan. But hey, I bit the bullet and the First Aid is currently running (checking 7 / 25 TimeMachine snapshots for some reason currently). The roof hasn't collapsed yet.
Even with that hugely fast SSD though, it's not moving that fast during this process.
Well, that took a while and it did exactly nothing. `fsck` was really pissy and refused to run as this forum thread (and others) instructed. I tried many varieties, `fsck_apfs` included, and in both cases had to specify the exact disk every time, otherwise it just checked a very small image that was mounted and currently the root filesystem (I am guessing the recovery mode image). I used Disk Utility's "First Aid" to double-check the right disk name and tried to run both fsck commands with several different options but was always stopped by an error message that the disk is mounted with write access and repairs are not possible. Oh well.
Searched the net for a bit and found a couple of StackOverflow posts (this one and that one) that basically say that this is no longer possible to do in Catalina as it was before it. Sigh. In any case, I've had a successful First Aid finish during the Single User mode but not sure it actually changed anything on the disk (and it did report a few inconsistencies during its checks of 25 TimeMachine snapshots).
At this point, and for now, I am just giving up. I know Apple is hellbent on protecting most of the users from shooting themselves (by acting like professionals when they clearly aren't) but you should be able to do something as basic as `fsck` during a single user mode, damn it.
Searched the net for a bit and found a couple of StackOverflow posts (this one and that one) that basically say that this is no longer possible to do in Catalina as it was before it. Sigh. In any case, I've had a successful First Aid finish during the Single User mode but not sure it actually changed anything on the disk (and it did report a few inconsistencies during its checks of 25 TimeMachine snapshots).
At this point, and for now, I am just giving up. I know Apple is hellbent on protecting most of the users from shooting themselves (by acting like professionals when they clearly aren't) but you should be able to do something as basic as `fsck` during a single user mode, damn it.
Try :
Boot into recovery (or a separate boot drive ) and get "off" the file system. Run "first Aid" on that.
On some occasions hve seen improvements in that context. APFS seems to have a narrow set of corner case issues of fixes while it is being actively used ( and vm/page variant is running).
If you want to try fsck directly can get to terminal in recovery mode also ( just have to pass the file system target to the command. )