Performance Tuning

Portworx has best practices for both global container level optimization, as well as volume granular optimization.

Global performance tuning

As of PX version 1.3, it is recommended to use a journal device to absorb PX metadata writes. Journal writes are small with frequent syncs and therefore only SSD/NVME should be configured as a journal device.

The journal device should be 2GB. Using a larger device will not help, since PX will only use 2GB of the journal device. The journal device can be specified via the -j option during installation, documented here.

If you are upgrading to 1.3 and want to add a journal device to an existing node, follow these instructions.

Volume granular performance tuning

By default, PX will try to auto tune the IO profile setting for a given volume by learning from the access patterns. However, this algorithm can be overridden and a specific profile can be chosen.

The IO profile can be selected while creating the volume via the io_profile flag. For example:

pxctl volume create --size=10 --repl=3 --io_profile=sequential demovolume

or

docker volume create -d pxd io_priority=high,size=10G,repl=3,io_profile=random,name=demovolume

It is highly recommended to let PX decide the correct IO profile tuning. If you do however override the setting, you should understand the operation of each profile setting.

Sequential

This optimizes the read ahead algorithm for sequential access. Use io_profile=sequential.

Random

This records the IO pattern of recent access and optimizes the read ahead and data layout algorithms for short term random patterns. Use io_profile=random.

CMS

This is useful for content management systems, like WordPress. This option applies to a PX shared (global namespace) volume. It implements an attribute cache and supports async writes. This increases the PX memory footprint by 100MB. Use io_profile=cms.

DB

This implements a write-back flush coalescing algorithm. This algorithm attempts to coalesce multiple syncs that occur within a 50ms window into a single sync. Coalesced syncs are acknowledged only after copying to all replicas. In order to do this, the algorithm requires a minimum replication (HA factor) of 3. This mode assumes all replicas do not fail (kernel panic or power loss) simultaneously in a 50 ms window. Use io_profile=db.