Share this post

April 01, 2016

Sysbench OLTP Performance: Atlantis HyperScale and Intel® Xeon® E5-2600 v4 CPUs

Hugo Phan - Atlantis Computing

Avatar

As part of our commitment to making available the latest processor models on Atlantis HyperScale hyperconverged appliances, we were able to test Intel’s newly released Xeon® E5-2600 v4 “Broadwell-EP” based processors starting with the Sysbench OLTP MySQL database. We took an existing HyperScale CX-12 that ships with Intel Xeon E5-2600 v3 “Haswell” processors and replaced the CPUs with Intel Xeon E5-2697A v4 processor, which are 2.6GHz (16C/32C) processor. With 16 cores of compute power per CPU and 8 CPUs in a 4-node 2 Rack Unit Atlantis HyperScale CX-12, we have 128 CPU cores at our disposal. With this impressive amount of processing power, we were able to achieve 2783 transaction per second on a single appliance. To put this in perspective, Wikipedia uses MySQl and has peak database transactions of 25,000 transactions per second (TPS)[i]. This means that nine Atlantis HyperScale CX-12 appliances could handle the transaction load of the entire Wikipedia database.[ii]

Sysbench OLTP Benchmark

Sysbench is a system performance benchmark that includes an Online Transaction Processing (OLTP) test profile. The OLTP test is not an approximation of an OLTP test, but is rather a true database-backed benchmark that conducts transactional queries to an instance of MySQL in a CentOS environment.

The first step in setting up the benchmark is to create the database itself, which is done by specifying the number of tables as well as the number of rows per table. In our test, we defined 100 tables with 100 million rows each, which resulted in a database of 1 billion entries. This database was 260GB.

To measure the performance of the Atlantis HyperScale cluster in a transactional database workload, we first leveraged the Sysbench OLTP benchmark, paying close attention to total aggregate performance. The Sysbench OLTP benchmark runs on top of Percona MySQL leveraging the InnoDB storage engine operating inside a CentOS installation. While a traditional SAN infrastructure can better cope with large single workloads, hyperconverged systems are designed to spread that load across all nodes in the system. To that end, we deployed four Sysbench VMs on the cluster, 1 per node, and measured the total performance seen on the cluster with all operating simultaneously.

Hardware Configuration

Four SuperMicro SuperServer SYS-2028TP-HC1R, each of the 4 nodes has:

  • Two Intel® Xeon® E5-2697A v4 2.6GHz (16/32)
  • 384GB Memory
  • Three Intel® DC S3510 800GB SSD
  • Intel® 82599EB 10GbE network adaptor

Each Sysbench VM is configured with two virtual disks, one for boot (40GB), one for the database that we will test (400GB). From a system resource perspective, we configured each VM with 16 vCPUs, 64GB RAM, the LSI Logic SAS SCSI controller and VMware Tools. It should be stressed that this configuration wasn’t designed to completely saturate all resources in the HyperScale cluster and in fact it left many resources free.

Sysbench Testing Configuration (per VM)

  • CentOS 6.4 64-bit
  • Percona XtraDB 5.5.30-rel30.1
    • Database Tables: 100
    • Database Size: 10,000,000
    • Database Threads: 32
    • RAM Buffer: 24GB
  • Test Length: 12 hours
    • 6 hours pre-conditioning 32 threads
    • 1 hour 32 threads
    • 1 hour 16 threads
    • 1 hour 8 threads
    • 1 hour 4 threads
    • 1 hour 2 threads

Results Summary

With four VMs operating simultaneously across the cluster, we measured peak 32-thread individual VM performance of 701.43 TPS, 695.45 TPS, 695.54 TPS and 690.32 TPS across the hosts. This gave us an average of 695.83 TPS from all four VMs, with both the slowest and the fastest being around 1% variation from the average. In total we measured an aggregate 2,783.32 TPS across the HyperScale cluster with four Sysbench VMs running.

Looking at the average latency in the Sysbench hyper-converged test, we saw response times measuring 45.62ms, 46.01ms, 46.01ms and 46.31ms under full load. Average across the entire cluster came in at 45.98ms.

HyperScale Node Transactions Per Second
sysbench1 node 701.43
sysbench2 node 695.45
sysbench3 node 695.54
sysbench4 node 690.90
Aggregate 2783.32

Atlantis HyperScale Processor Availability

This test was performed with pre-release Intel Xeon E5-2697A v4 samples provided by Intel that are not currently available in Atlantis HyperScale appliances. Atlantis is actively working with Intel and our server partners to add the v4 processors to the Atlantis HyperScale product line. The specific CPU choices in the general availability release of Atlantis HyperScale have not yet been determined and may differ from the tested configuration. Atlantis will continue to conduct and publish additional performance benchmark testing using Intel Xeon E5-2600 v4 processors. Stay tuned for more results coming at http://blog.atlantiscomputing.com


[ii] This is just an example to put the transactions per second achieved by Atlantis HyperScale in perspective. Wikipedia does not use Atlantis HyperScale for its MySQL database. 
12345
Current rating: 4.2 (5 ratings)