Thursday Sep 29, 2011

SPARC T4 Processor Outperforms IBM POWER7 and Intel (Westmere AES-NI) on OpenSSL AES Encryption Test

Oracle's SPARC T4 processor is faster than the Intel Xeon X5690 (with AES-NI) and the IBM POWER7.

  • On single-thread OpenSSL encryption, the 2.85 GHz SPARC T4 processor is 4.3 times faster than the 3.5 GHz IBM POWER7 processor.

  • On single-thread OpenSSL encryption, the 2.85 GHz SPARC T4 processor is 17% faster than the 3.46 GHz Intel Xeon X5690 processor.

The SPARC T4 processor has Encryption Instruction Accelerators for encryption and decryption for AES and many other ciphers. The Intel Xeon X5690 processor has AES-NI instructions which accelerate only AES ciphers. The IBM POWER7 does not have cryptographic instructions, but cryptographic coprocessors are available.

Performance Landscape

The table below shows results when running the OpenSSL speed command with the AES-256-CBC cipher. The reported results are for a message size of 8192 bytes. Results are reported for a single thread and for running on all available hardware threads (no over subscribing).

OpenSSL Performance with
AES-256-CBC Encryption
Processor Performance (MB/sec)
1 Thread Maximum Throughput
(at number of threads)
SPARC T4, 2.85 GHz 769 11,967 (64)
Intel Xeon X5690, 3.46 GHz 660 7,362 (12)
IBM POWER7, 3.5 GHz 179 2,860 (est*)

(est*) The performance of the IBM POWER7 is estimated at 16 times the rate of the single thread performance. The estimate is considered an upper bound on expected performance for this processor.

Configuration Summary

SPARC Configuration:

SPARC T4-1 server
1 x SPARC T4 processors, 2.85 GHz
64 GB memory
Oracle Solaris 11

Intel Configuration:

Sun Fire X4270 M2 server
1 x Intel Xeon X5690 processors, 3.46 GHz
24 GB memory
Oracle Solaris 11

Software Configuration:

OpenSSL 1.0.0.d
gcc 3.4.3

Benchmark Description

The in-memory SSL performance was measured with the openssl command. openssl has an option for measuring the speed of various ciphers and message sizes. The actual command used to measure the speed of AES-256-CBC was:

openssl speed -multi {number of threads} -evp aes-256-cbc

openssl runs for several minutes and measures the speed, in units of MB/sec, of the specified cipher for messages of sizes 16 bytes to 8192 bytes.

Key Points and Best Practices

  • The Encryption Instruction Accelerators are accessed through a platform independent API for cryptographic engines.
  • The OpenSSL libraries use the API. The default is to not use the Encryption Instruction Accelerators.
  • Cryptography is compute intensive. Using all available threads streams, both the SPARC T4 processor and the Intel Xeon processor were able to saturate the memory bandwidth of the respective systems.

See Also

Disclosure Statement

Copyright 2011, 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 9/26/2011.

SPARC T4-2 Server Beats Intel (Westmere AES-NI) on SSL Network Tests

Oracle's SPARC T4 processor is faster and more efficient than the Intel Xeon X5690 processor (with AES-NI) when running network SSL thoughput tests.

  • The SPARC T4 processor at 2.85 GHz is 20% faster than the 3.46 GHz Intel Xeon X5690 processor on single stream network SSL encryption.

  • The SPARC T4 processor requires fewer streams to attain near-linespeed of a 10 GbE secure network and does this with 5 times less CPU resources compared to the Intel Xeon X5690 processor.

  • Oracle's SPARC T4-2 server using 8 threads achieves line speed over a 10 GbE network with only 9% CPU utilization.

  • Oracle's Sun Fire X4270 M2 with two Intel Xeon X5690 processors achieves line speed with 8 threads, but at 45% CPU utilization.

The SPARC T4 processor has hardware support via Encryption Instruction Accelerators for encryption and decryption for AES and many other ciphers. The Intel Xeon X5690 processor has AES-NI instructions which accelerate only AES ciphers.

Performance Landscape

The following table shows single stream results running encrypted (SSL Read) and unencrypted (Clear Text) messages of 1 MB in size. These tests were run with the uperf benchmark and used the AES-256-CBC cipher. They were run across a 10 GbE connection. Write messages saw similar performance.

Single Stream Network Communication with Uperf
Processor Performance (Mb/sec)
Clear Text SSL Read
SPARC T4, 2.85 GHz 4,194 1,678
Intel Xeon X5690, 3.46 GHz 5,591 1,398

The next table shows how many streams it takes to achieve 90% of the 10 GbE network bandwidth (9000 Mb/sec) for encrypted read messages of 1 MB in size. These tests were run with the uperf benchmark and used the AES-256-CBC cipher. Write messages saw similar performance.

Uperf SSL Read with AES-256-CBC
Processor Number of
Streams for 90%
Network Utilization
CPU Utilization
SPARC T4, 2.85 GHz 8 9%
Intel Xeon X5690, 3.46 GHz 12 45%

Configuration Summary

SPARC T4 Configuration:

2 x SPARC T4-2 servers each with
2 x SPARC T4 processors, 2.85 GHz
128 GB memory
1 x 10-Gigabit Ethernet XAUI Adapter
Oracle Solaris 11
Back-to-back 10 GbE connection

Intel Configuration:

2 x Sun Fire X4270 M2 servers each with
2 x Intel Xeon X5690 processors, 3.46 GHz
48 GB memory
1 x Sun Dual Port 10GbE PCIe 2.0 Networking Card with Intel 82599 10GbE Controller
Oracle Solaris 11
Back-to-back 10 GbE connection

Software Configuration:

OpenSSL 1.0.0.d
uperf 1.0.3
gcc 3.4.3

Benchmark Description

Uperf is an open source benchmark program for simulating and measuring network performance. Uperf is able to measure the performance of various protocols, including TCP, UDP, SCTP and SSL. The uperf benchmark uses an input-defined workload to test network performance. This input workload can be used to model complex situations or to isolate simple tasks. The workload used for these tests was simple network reads and simple network writes.

Key Points and Best Practices

  • The Encryption Instruction Accelerators are accessed through a platform independent API for cryptographic engines.
  • The OpenSSL libraries use the API. The default is to not use the Encryption Instruction Accelerators.
  • Cryptography is compute intensive. Using 8 streams, the SPARC T4 processor was able to match the bandwidth of the 10 GbE network with 8 threads.

See Also

Disclosure Statement

Copyright 2011, 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 9/26/2011.

About

BestPerf is the source of Oracle performance expertise. In this blog, Oracle's Strategic Applications Engineering group explores Oracle's performance results and shares best practices learned from working on Enterprise-wide Applications.

Index Pages
Search

Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today