SAS Writeback cache performance & Device vs file extent performance disparities

[Stux]

I would suggest running the tests with sync=always|standard|disabled

It might be that the default for a file extent is different to a zvol extent for instance.

Also, you might find ZFS is actually bottlenecked by either single-thread CPU performance, or Memory bandwidth. Haven't determined yet, but I see roughly twice the speed when benchmarking my 2.2Ghz dual channel Xeon, vs my 4ghz quad channel.

One peaks at circa 1GB/s

(same generation)

Did not look into file extents too much, but as the file extents are repeatedly mentioned as "just for dev testing", then it wouldn't surprise me if data reliability shortcuts are taken (for example, not actually sync writing data that needs to be sync written)

 

[tvsjr]

Did not look into file extents too much, but as the file extents are repeatedly mentioned as "just for dev testing", then it wouldn't surprise me if data reliability shortcuts are taken (for example, not actually sync writing data that needs to be sync written)

Eww, that would be very ugly. In that case, the option shouldn't even be presented to a user in a final/production build.

 

[Chris Moore]

Eww, that would be very ugly. In that case, the option shouldn't even be presented to a user in a final/production build.

Since it is open source, there is probably not a difference between a production and a developer build other than the developer build might include programming tools.

There are people who download the source code and build it for themselves with their own tweaks along the way.

Sent from my SAMSUNG-SGH-I537 using Tapatalk

 

[Ender117]

Hi OP, I know this is an old thread but I found myself doing similar things recently: https://forums.freenas.org/index.php?threads/on-disk-cache-and-zfs-performance.70267/.

I was doing single disk pools but did found out sync setting have major impact on performance. Have you done any additional tests recently?

 

[pro lamer]

device extents to be WAY slower than file extents

Noob's guess coming: maybe device extents are IOPS oriented? Maybe you can make more 4K blocks comparison tests?

Sent from my mobile phone