enterprisestorage.com

The Era of Legacy PCIe SSDs is Over — A Single Micron 9100MAX NVMe SSD Drives Performance and Value in a PostgreSQL OLTP Workload

9100MAX vs. Legacy PCIe SSD Fast FactsNonvolatile memory express (NVMe) SSDs promise both dramatic performance improvements and the path forward for new data center build outs when compared to legacy application host controller interface (AHCI) PCI Express SSDs.

NVMe moves fast storage closer to the platform's processors to improve overall storage throughput and reduce application response time.

In this technical brief, we compare the future — a cutting edge Micron 9100MAX NVMe SSD with the past — a legacy PCIe SSD using PostgreSQL with an OLTP workload.

PostgreSQL is a widely deployed enterprise-class, open-source, object-oriented relational database management system that offers outstanding reliability and robust, consistent data integrity. It supports a variety of workloads in industries ranging from bio/pharmaceutical, healthcare, education, finance, e-commerce, media, manufacturing and telecommunications. One of the most demanding workloads it supports is OLTP, which is often used to manage transaction-based applications like order entry and fulfillment, real-time data acquisition, management and analysis, and large-scale commercial processes — all of which require immediate access to business-critical data. To support these real-time OLTP applications, platforms require extremely fast transaction processing and ultra-low latency.

Accelerating transaction processing yields better outcomes; additional transactions can be executed and managed within the same timeframe. Reducing storage response times enables the relational database management system to be more responsive to user requests so users and/or automation systems are not idle as they wait for storage I/O processes to complete.

9100MAX: 69% Better Throughput

Figure 1: New Orders per Minute - Legacy PCIe SSD (1) vs. 9100MAX SSD (1)New orders per minute (NOPM) is a good measure of database performance for OLTP and can be used to compare business throughput across different implementations.

The 9100MAX demonstrated as much as 69% better OLTP throughput (with 64 users) than the legacy PCIe SSD, as shown in Figure 1 (taller is better). A similar trend can be seen across user counts from 8 to 64, with NOPM improvements ranging from about 6% (8 users) up to the measured maximum.

9100MAX: 69% More Transactions

Figure 2: Transactions per Minute - Legacy PCIe SSD (1) vs. 9100MAX SSD (1)Transactions per minute (TPM) is a slightly different performance metric than NOPM. TPM measures database commits — enabling a direct performance comparison (if the database configuration and number of warehouses is kept constant, as it was in this example).

Figure 2 shows that a single 9100MAX was able to reach as much as 69% better TPM than the legacy PCIe SSD, with a similar trend ranging from about 6% better through the measured maximum (taller is better).

9100MAX: 41% Better Average, 20% Better 99.9th Percentile Response Times

Figure 3: Average Response Time (Latency) - Legacy PCIe SSD (1) vs. 9100MAX SSD (1)While NOPM and TPM express database throughput, average response time shows how quickly users and applications can interact with the database, and the 99th percentile response time shows how consistent that interaction is. In both cases, lower is better.

Figure 3 shows average database response time for an increasing number of users — from 8 to 64 — lower is better.

Consistently, latency is as important for many OLTP use models as low-average latency. Latency consistency can be characterized by looking at the 99th percentile, and the overall consistency can then be easily compared between storage types such as NVMe SSDs and legacy PCIe SSDs.

Figure 4: 99.9th Percentile Response Time (Latency) - Legacy PCIe SSD (1) vs. 9100MAX SSD (1)Figure 4 shows the 99.9th percentile latency differences between the 9100MAX and legacy PCIe SSDs as a function of increasing user count (from 8 to 64) — lower is better.

Figure 4 shows that when less than 32 users are active, the legacy PCIe SSD offers a lower (better) 99.9th percentile latency. However, as the database gets more active and the number of users rises (from 32 on), the 9100MAX shows consistently lower 99.9th percentile latency.

The Bottom Line

This technical brief shows clear differences — in both raw performance and performance consistency — between the Micron 9100MAX NVMe SSD and a legacy PCIe SSD.

Using a PostgreSQL database and an OLTP workload, a single Micron 9100MAX NVMe SSD delivers on the promise of NVMe:

  • New orders per minute: Up to 69% better
  • Transactions per minute: Up to 69% better
  • Average response time: As much as 41% lower
  • Response time consistency: As much as 20% better

The 9100MAX moves OLTP workloads forward, supporting the most demanding requirements and ushering them into next generation data management — all at the speed of now.

How We Tested

Test Database and Schema

To ensure consistent maximum performance from the PostgreSQL database, some PostgreSQL parameters were modified from their default values (using PgTune), as shown in Table 1.

Parameter Value
Default_statistics_target 100
Maintenance_work_mem 2GB
Effective_cache_size 72GB
Work_mem 768MB
Shared_buffers 30GB
Checkpoint_segments 4096
Checkpoint_timeout 5 min
Checkpoint_completion_target 0.8
Seq_page_cost 0.5
Random_page_cost 2.0
Bgwriter_delay 15
Bgwriter_Iru_maxpages 1000
Effective_io_concurrency 1000

Table 1: PostgreSQL Database Parameters

For testing, a warehouse/district/customer/item model was used, composed of an approximately 800GB database (larger than the available memory), resulting in a read-heavy storage I/O workload. For each test, the schema was identical to the tables and row counts shown in Table 2.

Table Number of Rows
NEW_ORDER 67,500,000;
ORDER_LINE 2,250,000,000
CUSTOMER 225,000,000
HISTORY 75,000
ITEM 750,000
WAREHOUSE 7,500
ORDER 2,250,000,000
STOCK 750,000,000

Table 2: Schema Tables and Row Count

Platform Configuration

To ensure consistent and comparable results, identical hardware and software platforms and databases were tested. (Note: Database settings and file locations changed as part of the testing.) For each test, the test sequence was identical.

Database Server

  • Dell™ PowerEdge™ R730xd
  • 2 Intel® Xeon® E5-2690 v3 processors
  • 128GB RAM
  • 1x 1.2TB Micron 9100MAX NVMe SSD
  • 1x 1.4TB legacy (AHCI) SSD with optimized driver loaded
  • CentOS7
  • PostgreSQL 9.4.6
  • Linux® NVMe v1.0

Test Sequence

The test sequence for the 9100MAX and legacy PCIe SSDs was identical. The SSDs were securely erased prior to use (to restore to a factory-fresh state), then the XFS file system was created (also with the default options), the device was mounted, and permissions were set for the user. The PostgreSQL metadata was moved to the new volume, PostgreSQL was started, and the test run started. For each test there was a ramp-up time (45 minutes) followed by performance measurement for 20 minutes, after which the data was collected and plotted.

SSD Details

In this series of tests, the 9100MAX was matched as closely as possible to a very high-performance legacy PCIe SSD. Details are shown in table 3.

Specification 9100MAX SSD Legacy AHCI SSD
Interface PCIe Gen3 PCIe Gen2
Lane Count 4 8
Bandwidth1 8 GB/s 8 GB/s
Capacity2 1.2TB 1.4TB;
Form factor Card Card

Table 3: SSD Specifications

1 Bandwidth is the theoretical maximum based on the PCIe interface specification.
2 A larger capacity SSD generally offers better performance.


This technical marketing brief is published by Micron and has not been authorized, sponsored, or otherwise approved by The PostgreSQL Global Development Group. Products are warranted only to meet Micron’s production data sheet specifications. Products and specifications are subject to change without notice. ©2016 Micron Technology, Inc. Micron and the Micron logo are trademarks of Micron Technology, Inc. PostgreSQL is copyright © 1996-8 by the PostgreSQL Global Development Group, and is distributed under the terms of the Berkeley license. NVMe is a trademark of NVM Express, Inc. PCIe is a registered trademark of PCI-SIG. Dell is a trademark of Dell Inc. Intel and Xeon are trademarks of Intel Corporation or its subsidiaries in the U.S. and/or other countries. CentOS is a trademark of Red Hat, Inc. Linux is a registered trademark of the Linux Foundation. All other trademarks are the property of their respective owners. All rights reserved. Rev. A 11/16, CCMMD-676576390-10399