Working Principles

I/O Traffic Control

I/O traffic control is implemented by LUN-based I/O queue management, token allocation, and dequeue management.

LUN-based I/O queue management uses a token mechanism to allocate storage resources. SmartQoS determines the number of storage resources to be allocated to an I/O queue of an LUN by counting the number of tokens owned by the queue. The more tokens owned by an I/O queue, the more resources will be allocated to that queue. Accordingly, I/Os in that queue will be preferentially processed. Figure 1-1 shows how SmartQoS manages I/O queues.

Figure 1-1 Managing I/O queues

The process of managing I/O queues is described as follows:

1. After application servers send I/O requests, the storage system delivers the requests to corresponding I/O queues.
2. The storage system adjusts the number of tokens owned by LUN I/O queues based on the priorities of LUNs. By reducing the number of tokens owned by queues that have a low priority, the storage system ensures that sufficient resources are available for high-priority LUNs so that I/Os to these LUNs can be preferentially processed.
3. The storage system processes the I/Os in queues by priority.

Example 1: Traffic control of LUNs in different SmartQoS policies

Hierarchical Management

SmartQoS supports both common and hierarchical policies.

• Common policy: controls the traffic from a single type of application to LUNs.
• Hierarchical policy: controls the traffic from various types of applications to LUNs. Common policies can be added to hierarchical policies.

Figure 1-2 shows their relationship.

Figure 1-2 Relationship between common and hierarchical policies

Burst Traffic Control Management

For latency-sensitive services, you can allow them to exceed the upper limit for a specific period of time. SmartQoS supports burst traffic control management to specify the burst IOPS, bandwidth, and duration for LUNs, LUN groups, or hosts.

The system accumulates the unused resources during off-peak hours and consumes them during traffic bursts to break the upper limit for a short period of time. To achieve this, the long-term average traffic of the service should be below the upper limit.

Minimum Performance Assurance

SmartQoS suppresses the performance of LUNs that have not been configured with the minimum performance assurance policy to guarantee the performance of LUNs that have been configured with this policy. The system implements minimum performance assurance as follows:

1. Allocates the minimum performance assurance objectives configured in the SmartQoS policy to each LUN in the policy based on performance requirements.
2. Sets the default minimum performance assurance objective for LUNs that are not added to the minimum performance assurance SmartQoS policy to ensure that services on the LUNs are not interrupted.
3. Calculates the performance gap and margin based on the minimum performance assurance objective and current performance of each LUN.
   • The value of the minimum performance assurance objective exceeding the actual performance is the performance margin.
   • The value of the actual performance exceeding the minimum performance assurance objective is the performance gap.
4. Suppresses the performance of the top 10 LUNs with the largest performance margin until no performance gap exists in the system.
   • In scenarios where the service pressure is unstable (that is, the overall IOPS and bandwidth of services fluctuate dramatically), it is not advised to configure any minimum performance assurance policy. With unstable service pressure, SmartQoS re-adjusts the resource allocation between services with the minimum performance assurance policy and services without such a policy, which may cause the service performance fluctuate more dramatically.
   • If the minimum performance assurance policy is configured, SmartQoS will limit resources allocated to services to which the policy is not applicable. As a result, the performance of these services will deteriorate and latency will increase. It is advised to evaluate whether the impact is acceptable before setting the minimum performance assurance policy.

Objective Distribution

All objects in a SmartQoS policy share the upper limit objectives. SmartQoS periodically collects the performance and requirement statistics of all LUNs or hosts in a traffic control policy, and distributes the traffic control objective to each LUN or host.

Currently, the optimized weighted max-min fairness algorithm is used for objective distribution. It determines the traffic control objective for each object (that is, LUN or host) based on the policy's overall objective and each object's resource requirement. The algorithm preferentially meets each object's resource requirement, and then distributes remaining resources to each object based on the object's weight. In addition, it uses a filtering mechanism to ensure a relative stable objective for each object.