Service Levels and their target response times

The following table lists the service level prioritizations and their associated performance target goals. When using SLs, critical systems can be assigned a high SL such as Diamond, Platinum, or Gold to ensure that their performance goals have a higher priority over applications with a lower SL such as Silver, Bronze, or Optimized.

In addition, some SLs have a lower response time limit, meaning that assigning them to SGs will force read and write latencies no lower than that limit. Also, notice that Diamond and Platinum SLs have lower target response times when SCM drives are used.

Table 4.              Service Levels priorities and limits

Unlike any of the other SLs, Optimized does not have a specific performance target. If SGs with a non-Optimized SL struggle to maintain their performance goals, they can add latency to SGs with Optimized SL as they try to preserve their own goals.

Similarly, if any non-Optimized SGs struggle to maintain their performance goals, they can add latency to SGs that are set with a lower priority SL. For example, Diamond SGs can affect Platinum SGs, which can affect Gold SGs, and so on.

Service Levels and SCM drives

The PowerMaxOS Q3 2019 release introduced PowerMax support for storage class memory (SCM), in addition to the existing NAND SSD flash drives. The support for different drive technologies introduces tiered storage called automated data placement (ADP). The functionality of ADP is included as part of the service levels (SLs) and uses machine learning to place data on the appropriate drive type. To improve the response times managed by SLs, ADP places the most active data on the faster drive technology.

ADP movements occur as either a promotion of data or a demotion of data. Promotions are the movement of data to the SCM drives, and demotions are the movement of data out of the SCM drives. PowerMaxOS uses the SL prioritization together with activity metrics to determine promotions and demotions, where the basic strategy is described in ‎Table 5.

Table 5.              SCM drives promotion and demotion strategy based on Service Levels

For more information about PowerMax Service Levels see: Dell EMC PowerMax: Service Levels for PowerMaxOS Technical White Paper.

Service Levels example with a single database workload

The following figure shows the typical effect of service levels on a single Oracle database workload. In this test, a single OLTP workload ran without interruption or change. Only the data_sg SL changed every 30 minutes.

Figure 23.               Service Level changes on a single Oracle workload

We see that a Bronze SL forced an average of 5 ms latency, and a performance level of 37,000 IOPS. After the SL changed to Silver, the latency dropped to 2 ms and IOPS increased to 79,000. The Gold SL reduced latency to 0.4 ms and IOPS increased to 192,000. The Platinum and Diamond SLs were not significantly different as they both performed at 0.3 ms latency and 204,000 IOPS.

When a SL changes, the effect is immediate because it takes place at the PowerMaxOS software layer. We also see that SL latencies affect both reads and writes I/O response times.

Service Levels example with two database workloads

The following figure shows the effect of Service Levels on two Oracle databases. In this test, two OLTP workloads ran without interruption or change. As seen on the left of the figure, the SL of the workload (represented by the top yellow line) is set to Diamond SL simulating a key mission critical application. The SL of the other workload (represented by the lower blue line) started at Bronze SL and was increased every 30 min until it reached Diamond SL.

Figure 24.               Service Level changes on two Oracle workloads

We can see that as the “caged” application improved its SL, it slowly took more resources from the application with the Diamond SL, until they shared the same SL and system resources. This result demonstrates the value of setting a lower priority SL to lower priority applications.