EXT4 File-System Tuning Benchmarks

[Flynn]

I would have liked to see the performance impact of the options atime, relatime and noatime.

[jwilliams]

actually your question gets answered if you look on the very first page of the article. there you find the exact paramters used for each test. "discard" meand with discard/trim activated. that sentence was a mistake.

Actually, I read the first page, and actually, no, it does not say that it means "with discard/trim activated".

[a user]

I thought giving the discard option enables TRIM. But you wrote that disabling TRIM slowed things down in one test.

Did you mean that enabling TRIM (i.e., giving the discard mount option) slows down the Dbench v4.0 test?

actually your question gets answered if you look on the very first page of the article. there you find the exact paramters used for each test. "discard" meand with discard/trim activated. that sentence was a mistake.

[a user]

This test also misses changing i/o the scheduler. Using the noop scheduler produces better performance for SSDs and hardware RAIDs.

no, not true!

actually deadline is in general the best io scheduler for ssds, though it may need some tweaking to get to full power. but even on defaults its far superior to noop. there are only rare cases where noop can get slightly the tip over deadline.

in my own tests on 4 different ssds i never had a single time noop won any of the tests. actually i never won any of the tests though it came out on second place a few times. deadline was the best overall. same seem to confirm a lot of other people and tests.

[jwilliams]

I thought giving the discard option enables TRIM. But you wrote that disabling TRIM slowed things down in one test.

Did you mean that enabling TRIM (i.e., giving the discard mount option) slows down the Dbench v4.0 test?

[phred14]

Once upon a time, I came upon the advice that for an NFS server one could get better write performance at the client with "data=journal" - and an over-sized journal. Reason being that the NFS server was ready to call the client write complete as soon as the journal was written, rather than waiting for the full write to complete.

[Artemis3]

This test also misses changing i/o the scheduler. Using the noop scheduler produces better performance for SSDs and hardware RAIDs.

[roland]

How does data=journal actually improve performance on some of these? Anway, thanks for the great defaults ext devs!

[mayankleoboy1]

Y U no HDD?

Why dont you do the same tests on a HDD as well ? They are far more common than a SSD.
Also, a SSD owner would already be getting far more performance compared to a HDD owner. The onus is on the HDD owner to try to get as much speed as safely possible.

[Zplay]

Michael, could you do a benchmark of EXT4 with different journal size ?