DirectIO For OpenZFS Shows Very Promising Performance
Running the past two days was the annual OpenZFS Developer Summit. One of the most interesting presentations from this virtual event was on the status of DirectIO (O_DIRECT) support for the OpenZFS file-system and the performance boost it can offer in relevant areas.
Brian Atkinson of the Los Alamos National Laboratory presented at the developer summit around the DirectIO support for OpenZFS. The work ultimately boils down to the DirectIO merge request open for OpenZFS since February 2020. This support aims to allow bypassing ZFS' ARC when issuing reads/writes with a particular focus on improving the performance for Zpools on NVMe solid-state drives as well as other cases where ARC just gets in the way.
When using speedy NVMe drives as VDEVs, he found a 1.5~3x speed-up using the DirectIO mode.
The Linux support/performance at least appears to be in good shape while the FreeBSD performance testing is ongoing. The pull request remains open but could land potentially for OpenZFS 3.0.
Those interested in Brian's presentation on DirectIO for OpenZFS can see the slide deck.
Brian Atkinson of the Los Alamos National Laboratory presented at the developer summit around the DirectIO support for OpenZFS. The work ultimately boils down to the DirectIO merge request open for OpenZFS since February 2020. This support aims to allow bypassing ZFS' ARC when issuing reads/writes with a particular focus on improving the performance for Zpools on NVMe solid-state drives as well as other cases where ARC just gets in the way.
When using speedy NVMe drives as VDEVs, he found a 1.5~3x speed-up using the DirectIO mode.
The Linux support/performance at least appears to be in good shape while the FreeBSD performance testing is ongoing. The pull request remains open but could land potentially for OpenZFS 3.0.
Those interested in Brian's presentation on DirectIO for OpenZFS can see the slide deck.
12 Comments