Linux 3.12 Kernel Still Shows Disk Improvements

[phoronix]

Phoronix: Linux 3.12 Kernel Still Shows Disk Improvements

The Linux 3.12 kernel brings faster open-source AMD graphics due to a CPUfreq subsystem change but there also appears to be faster disk performance...

http://www.phoronix.com/vr.php?view=MTQ5MzI

[Ericg]

Any chance the disk improvements are also helped by the CPUfreq change?

[matyas]

Any chance the disk improvements are also helped by the CPUfreq change?

Nice one. This is what ondemand does to you when used for benchmarking. Hint: it shouldn't be used for benchmarking.

[helland]

Possible, but maybe not?

I don't know if this could be caused by the changes to the ondemand scheduler or not. My guess though, is that it's caused by fixing performance regressions. Looking back at http://www.phoronix.com/scan.php?pag...esystems&num=1 there seemed to be an overall regression for all filesystems from 3.9 to 3.10 to 3.11. The performance gained with 3.12 looks like it about matches the amount of regression on the two previous releases combined. That's just my 5 cents though.

[Figueiredo]

Any chance the disk improvements are also helped by the CPUfreq change?

I hope not. In my experience (ubuntu), disk access and transfer rates are painfully slow on linux, any improvement on this front will certainly be much appreciated.

[schmidtbag]

I hope not. In my experience (ubuntu), disk access and transfer rates are painfully slow on linux, any improvement on this front will certainly be much appreciated.

Linux disk benchmarks seem to vary pretty drastically. There are many times where a filesystem such as EXT4 is considerably faster than NTFS, and plenty of times where performance is half of what it should be. Overall I'd say linux filesystems are pretty good, but from what I heard, BSD has the best filesystems.

Supposing these benchmarks are not due to the ondemand governor, I look forward to these impressive new results. I have a 64GB SATA III SSD and I can't seem to get much higher than 230MB/s read speed - I should be getting double that. Write speeds seem to be pretty good, ranging from 300-450MB/s.

[Guest]

Linux disk benchmarks seem to vary pretty drastically. There are many times where a filesystem such as EXT4 is considerably faster than NTFS, and plenty of times where performance is half of what it should be. Overall I'd say linux filesystems are pretty good, but from what I heard, BSD has the best filesystems.

What I've seen Linux was always faster than ntfs, so show me benchmarks you're talking about. BSD have old UFS and experimental zfs. None of them is comparable to Linux file systems in any way (except zfs in term of features). And they're much slower, that's for sure.

[blackshard]

Some years ago (three maybe...) I did a empiric and very naive test on my notebook while transferring data from/to a ntfs partition.

I always considered ntfs-3g fuse driver kinda slow but on my machine it performed very slowly.
I noticed, using the ondemand governor, performances were nasty. I found that my cpu was not pushed to its highest frequency during transfers. Instead, using the conservative/performance governor, I had much better throughput.

[chrisb]

Linux disk benchmarks seem to vary pretty drastically. There are many times where a filesystem such as EXT4 is considerably faster than NTFS, and plenty of times where performance is half of what it should be. Overall I'd say linux filesystems are pretty good, but from what I heard, BSD has the best filesystems.

Supposing these benchmarks are not due to the ondemand governor, I look forward to these impressive new results. I have a 64GB SATA III SSD and I can't seem to get much higher than 230MB/s read speed - I should be getting double that. Write speeds seem to be pretty good, ranging from 300-450MB/s.

What benchmark did you use? And did you do it on a non-root partition mounted only for the benchmark? Peak read speed is quoted for bulk sequential access. Perhaps your reads are smaller, or non sequential, or other processes are also reading? Write speed may appear higher if writes are cached.

[schmidtbag]

What benchmark did you use? And did you do it on a non-root partition mounted only for the benchmark? Peak read speed is quoted for bulk sequential access. Perhaps your reads are smaller, or non sequential, or other processes are also reading? Write speed may appear higher if writes are cached.

I tried something as simple as copying large (hundred MB) files to and from a RAM disk. Nothing else was using the drive, the drive is properly calibrated, I have it attached to a SATA III port, and the partition is mounted by root (its the boot partition).