Tune performance in GKE

In its default configuration, Portworx attempts to provide good performance across a wide range of situations. However, you can improve your storage performance on your environment by configuring a number of settings and leveraging features Portworx offers. To get the most out of Portworx, follow the guidance provided in this article.

Configure the network data interface

You can provide Portworx with a specific network interface for data when you generate the spec as part of your installation. Portworx by Pure Storage recommends a network interface with a bandwidth of at least 10Gb/s and network latency below 5 milliseconds. If multiple NICs are present on the host, present a bonded interface to Portworx.

note

If you've already installed Portworx, you can update the network.dataInterface value of the install spec and reapply it.

Enable hyperconvergence

Use Stork to ensure your Pod is running on the same node in which the data resides.

Configure your cluster topology

When configured to be aware of your cluster topology, Portworx places replicas for high availability. Configure your cluster topology.

Define a VolumePlacementStrategy

StatefulSets, distributed NoSQL databases, such as Cassandra, require PVCs to be distributed across the cluster. Use Affinity/Anti Affinity rules along with topology labels to define relationships between PVCs.

Define a VolumePlacementStrategy using affinity and anti-affinity labels to distribute volumes.

Adjust storage classes

To improve performance, adjust storage class parameters in the following ways:

• Prioritize volume traffic by setting the priority_io: field to high

• Choose the replication factor best suited to your high availability needs

  kind: StorageClass
  apiVersion: storage.k8s.io/v1beta1
  metadata:
   name: px-storage-class
  provisioner: pxd.portworx.com
  allowVolumeExpansion: true
  parameters:
   repl: "2"
   priority_io: "high"
   nodiscard: "true"

Modify Portworx resource consumption

By default, Portworx consumes as little CPU and memory resources as possible. You can potentially improve performance by allocating more resources, allowing Portworx to use more CPU threads and memory. Do this by modifying the spec used to install Portworx based on your cluster architecture.

Disaggregated architecture

In disaggregated deployments with dedicated storage nodes, enable higher resource consumption by specifying the rt_ops_conf_high runtime option:

apiVersion: core.libopenstorage.org/v1
kind: StorageCluster
metadata:
  name: px-cluster
  namespace: <px-namespace>
spec:
  image: portworx/oci-monitor:3.1.6
  ...
  runtimeOptions:
    rt_opts_conf_high: "1"

Hyperconverged architecture

In a non-disaggregated/hyperconverged architecture, where applications are running on the same host as storage, set threads based on the number of cores that can be allocated to Portworx. For example, if your host has 16 cores:

• num_threads=16 sets the total number of threads performing storage operations.
• num_io_threads=12 sets the number of threads that can do IO operations out of the total num_threads. In general, IO threads should be 75% of total threads.
• num_cpu_threads=16 sets the number threads that can do operations other than IO out of the total num_threads.

Configure these values as runtime options:

apiVersion: core.libopenstorage.org/v1
kind: StorageCluster
metadata:
  name: px-cluster
  namespace: <px-namespace>
spec:
  image: portworx/oci-monitor:3.1.6
  ...
  runtimeOptions:
    rt_opts_conf_high: "1"
    num_threads: "16"
    num_io_threads: "12"
    num_cpu_threads: "16"

To see more StorageCluster examples, visit the StorageCluster section of the documentation.