X

Everything you want and need to know about Oracle SPARC systems performance

Yahoo Cloud Serving Benchmark: SPARC T7-4 With Oracle NoSQL Beats x86 E5 v3 Per Chip

Brian Whitney
Principal Software Engineer

Oracle's SPARC T7-4 server delivered 1.9 million ops/sec on 1.6 billion records for the Yahoo Cloud Serving Benchmark (YCSB) 95% read/5% update workload.  Oracle NoSQL Database was used in these tests. NoSQL is important for Big Data Analysis and for Cloud Computing.

  • One processor performance on the SPARC T7-4 server was 2.5 times better than one chip Intel Xeon E5-2699 v3 for the YCSB 95% read/5% update workload.

  • The SPARC T7-4 server showed low average latency of 1.12 msec on read and 4.90 msec on write while achieving nearly 1.9 million ops/sec.

  • The SPARC T7-4 server delivered 325K inserts/sec on 1.6 billion records with a low average latency of 2.65 msec.

  • One processor performance on the SPARC T7-4 server was over half a million (511K ops/sec) on 400 million records for the YCSB 95% read/5% update workload.

  • Near-linear scaling from 1 to 4 processors was 3.7x while maintaining low latency.

These results show the SPARC T7-4 server can handle a large database while achieving high throughput with low latency for cloud computing.

Performance Landscape

This table presents single chip results comparing the SPARC M7 processor (in a SPARC T7-4 server) to the Intel Xeon Processor E5-2699 v3 (in a 2-socket x86 server).  All of the following results were run as part of this benchmark effort.

Comparing Single Chip Performance on YCSB Benchmark
Processor Insert Mixed Load (95% Read/5% Update)
Throughput
ops/sec
Average Latency Throughput
ops/sec
Average Latency
Write msec Read msec Write msec
SPARC M7 89,617 2.40 510,824 1.07 3.80
E5-2699 v3 55,636 1.18 202,701 0.71 2.30
 

The following table shows the performance of the Yahoo Clouds Serving Benchmark on multiple processor counts on the SPARC T7-4 server.

SPARC T7-4 server running YCSB benchmark
CPUs Shards Insert Mixed Load (95% Read/5% Update)
Throughput
ops/sec
Average Latency Throughput
ops/sec
Average Latency
Write msec Read msec Write msec
4 16 325,167 2.65 1,890,394 1.12 4.90
3 12 251,051 2.57 1,428,813 1.12 4.68
2 8 170,963 2.52 968,146 1.11 4.37
1 4 89,617 2.40 510,824 1.07 3.80
 

Configuration Summary

System Under Test:

SPARC T7-4 server
4 x SPARC M7 processors (4.13 GHz)
2 TB memory (64 x 32 GB)
8 x Sun Storage 16 Gb Fibre Channel PCIe Universal FC HBA, Emulex
8 x Sun Dual Port 10 GbE PCIe 2.0 Low Profile Adapter, Base-T
 
Oracle Server X5-2L server
2 x Intel Xeon E5-2699 v3 processors (2.3 GHz)
384 GB memory
1 x Sun Storage 16 Gb Fibre Channel PCIe Universal FC HBA, Emulex
1 x Sun Dual 10GbE SFP+ PCIe 2.0 Low Profile Adapter
 

External Storage (Common Multiprotocol SCSI TARget, or COMSTAR enables system to be seen as a SCSI target device):

16 x Sun Server X3-2L servers
configured as COMSTAR nodes, each with
2 x Intel Xeon E5-2609 processors (2.4 GHz)
4 x Sun Flash Accelerator F40 PCIe Cards, 400 GB each
1 x 8 Gb dual port HBA

Please note: These devices are only used as storage. No NoSQL is run on these COMSTAR storage nodes. There is no query acceleration done on these COMSTAR storage nodes.
 

Software Configuration:

Oracle Solaris 11.3 (11.3.1.2.0)
Logical Domains Manager v3.3.0.0.17 (running on the SPARC T7-4)
Oracle NoSQL Database, Enterprise Edition 12c R1.3.2.5
Java(TM) SE Runtime Environment (build 1.8.0_60-b27)
 

Benchmark Description

The Yahoo Cloud Serving Benchmark (YCSB) is a performance benchmark for cloud database and their systems.  The benchmark documentation says:

With the many new serving databases available including Sherpa, BigTable, Azure and many more, it can be difficult to decide which system is right for your application, partially because the features differ between systems, and partially because there is not an easy way to compare the performance of one system versus another.  The goal of the Yahoo Cloud Serving Benchmark (YCSB) project is to develop a framework and common set of workloads for evaluating the performance of different "key-value" and "cloud" serving stores.

Key Points and Best Practices

  • The SPARC T7-4 server showed 3.7x scaling from 1 to 4 sockets while maintaining low latency.
  • Four Oracle VM for SPARC (LDom) servers were created per processor, for a total of sixteen LDoms.  Each LDom was configured with 120 GB memory accessing two PCIe IO slots under SR-IOV (Single Root IO Virtualization).

  • The Sun Flash Accelerator F40 PCIe Card demonstrated excellent IO capability and performed 841K read IOPS (3.5K IOPS per disk) during the 1.9 million ops/sec benchmark run.

  • There was no performance loss from Fibre Channel SR-IOV (Single Root IO Virtualization) compared to native.

  • Balanced memory bandwidth was delivered across all four processors achieving an average total of 304 GB/sec during 1.9 million ops/sec run.

  • The 1.6 billion records were loaded into 16 Shards with the replication factor set to 3.

  • Each LDom is associated with a processor set (16 total).  The default processor set was additionally used for OS and IO interrupts.  The processors sets were used to ensure a balanced load.

  • Fixed priority class was assigned to Oracle NoSQL Storage Node java processes.

  • The ZFS record size was set to 16K (default 128K) and this worked the best for 95% read/5% update workload.

  • A total of eight Sun Server X4-2 and Sun Server X4-2L systems were used as clients for generating the workload.

  • The LDoms and client systems were connected through a 10 GbE network.

See Also

Disclosure Statement

Copyright 2015, Oracle and/or its affiliates. All rights reserved. Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. Results as of Oct 25, 2015.

Be the first to comment

Comments ( 0 )
Please enter your name.Please provide a valid email address.Please enter a comment.CAPTCHA challenge response provided was incorrect. Please try again.Captcha
Oracle

Integrated Cloud Applications & Platform Services