This section shows how VMs can use PMEM in a scale-out fashion when traditional storage like SAN or SSDs become a bottleneck. We used Sysbench [15] which is a database performance benchmark against MySQL  database to demonstrate the scaling benefits of PMEM.

Highlights

  • 4 VMs scale to 3.5x with vPMEM compared to 2x with NVMe SSD
  • vPMEM throughput and latency are 5x compared to NVMe SSD in a single VM

Configuration

Table 9 gives the details of the VM.

OS

CentOS 7.4

CPU

12 vCPU

vRAM

16 GB

NVMe SSD

128 GB

vPMEMdisk

128 GB

vPMEM

128 GB

Table 9: Sysbench VM configuration

Table 10 gives the Sysbench parameters used.

DB Tables

6

Rows per Table

90 M

DB Size

120 GB

Sysbench Threads

12

Tests

oltp_read_only

oltp_read_write (75/25)

oltp_write_only

Table 10: Sysbench workload configuration

Single VM Results

Figure 12 shows the normalized throughput reported by Sysbench in different configurations. The vPMEM case yields up to 5.5x better throughput compared to NVMe SSD.

Figure 12: Sysbench throughput

 

Figure 13 shows the normalized application-level 95th percentile latency reported by Sysbench in different configurations. The vPMEM case yields 5x better latency compared to NVMe SSD.

Figure 13: Sysbench latency

 

Scale-Out Results

We used multiple copies (two and four) of the same VM from Table 7. Figure 14 shows the scaling of multiple VMs in different configurations for the oltp_read_write test. We observed that vPMEMDisk and vPMEM can scale to around 3.5x, while NVMe SSD stops at 2x. We see a similar trend in the other two tests (read_only and write_only).

Figure 14: Sysbench throughput scaling for oltp_read_write test

 

Figure 15 shows the latency scaling with multiple Sysbench VMs. We do not observe any significant increase in latency when multiple VMs are running on vPMEMDisk and vPMEM, while NVMe SSD suffers more than 2x latency when four VMs are running.

Figure 15: Sysbench latency scaling

check-circle-line exclamation-circle-line close-line
Scroll to top icon