X

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

SPECjbb2015: SPARC T7-1 World Record for 1 Chip Result

Brian Whitney
Principal Software Engineer

Updated November 30, 2015 to point to published results and add latest, best x86 two-chip result.

Oracle's SPARC T7-1 server, using Oracle Solaris and Oracle JDK, produced world record one-chip SPECjbb2015 benchmark (MultiJVM metric) results beating all previous one- and two-chip results in the process.  This benchmark was designed by the industry to showcase Java performance in the Enterprise. Performance is expressed in terms of two metrics, max-jOPS which is the maximum throughput number, and critical-jOPS which is critical throughput under service level agreements (SLAs).

  • The SPARC T7-1 server achieved 120,603 SPECjbb2015-MultiJVM max-jOPS and 60,280 SPECjbb2015-MultiJVM critical-jOPS on the SPECjbb2015 benchmark.

  • The one-chip SPARC T7-1 server delivered 2.5 times more max-jOPS performance per chip than the best two-chip result which was run on the Cisco UCS C220 M4 server using Intel v3 processors.  The SPARC T7-1 server also produced 4.3 times more critical-jOPS performance per chip compared to the Cisco UCS C220 M4.  The Cisco result enabled the COD BIOS option.

  • The SPARC T7-1 server delivered 2.7 times more max-jOPS performance per chip than the IBM Power S812LC using POWER8 chips.  The SPARC T7-1 server also produced 4.6 times more critical-jOPS performance per chip compared to the IBM server. The SPARC M7 processor also delivered 1.45 times more critical-jOPS performance per core than IBM POWER8 processor.
  • The one-chip SPARC T7-1 server delivered 3 times more max-jOPS performance per chip than the two-chip result on the Lenovo Flex System x240 M5 using Intel v3 processors.  The SPARC T7-1 server also produced 2.8 times more critical-jOPS performance per chip compared to the Lenovo.  The Lenovo result did not enable the COD BIOS option.

  • The SPARC T5-2 server achieved 80,889 SPECjbb2015-MultiJVM max-jOPS and 37,422 SPECjbb2015-MultiJVM critical-jOPS on the SPECjbb2015 benchmark.

  • The one-chip SPARC T7-1 server demonstrated a 3 times max-jOPS performance improvement per chip compared to the previous generation two-chip SPARC T5-2 server.

From SPEC's press release:

"The SPECjbb2015 benchmark is based on the usage model of a worldwide supermarket company with an IT infrastructure that handles a mix of point-of-sale requests, online purchases, and data-mining operations.  It exercises Java 7 and higher features, using the latest data formats (XML), communication using compression, and secure messaging."

The Cluster on Die (COD) mode is a BIOS setting that effectively splits the chip in half, making the operating system think it has twice as many chips as it does (in this case, four, 9 core chips).  Intel has said that COD is appropriate only for highly NUMA optimized workloads.  Dell has shown that there is a 3.7x slower bandwidth to the other half of the chip split by COD.

Performance Landscape

One- and two-chip results of SPECjbb2015 MultiJVM from www.spec.org as of November 30, 2015.

SPECjbb2015
One- and Two-Chip Results
System SPECjbb2015-MultiJVM OS JDK Notes
max-jOPS critical-jOPS
SPARC T7-1
1 x SPARC M7
(4.13 GHz, 1x 32core)
120,603 60,280 Oracle Solaris 11.3 8u66 -
Cisco UCS C220 M4
2 x Intel E5-2699 v3
(2.3 GHz, 2x 18core)
97,551 28,318 Red Hat 6.5 8u60 COD
Dell PowerEdge R730
2 x Intel E5-2699 v3
(2.3 GHz, 2x 18core)
94,903 29,033 SUSE 12 8u60 COD
Cisco UCS C220 M4
2 x Intel E5-2699 v3
(2.3 GHz, 2x 18core)
92,463 31,654 Red Hat 6.5 8u60 COD
Lenovo Flex System x240 M5
2 x Intel E5-2699 v3
(2.3 GHz, 2x 18core)
80,889 43,654 Red Hat 6.5 8u60 -
SPARC T5-2
2 x SPARC T5
(3.6 GHz, 2x 16core)
80,889 37,422 Oracle Solaris 11.2 8u66 -
Oracle Server X5-2L
2 x Intel E5-2699 v3
(2.3 GHz, 2x 18core)
76,773 26,458 Oracle Solaris 11.2 8u60 -
Sun Server X4-2
2 x Intel E5-2697 v2
(2.7 GHz, 2x 12core)
52,482 19,614 Oracle Solaris 11.1 8u60 -
HP ProLiant DL120 Gen9
1 x Intel Xeon E5-2699 v3
(2.3 GHz, 18core)
47,334 9,876 Red Hat 7.1 8u51 -
IBM Power S812LC
1 x POWER8
(2.92 GHz, 10core)
44,883 13,032 Ubuntu 14.04.3 J9 VM -


* Note COD: result uses non-default BIOS setting of Cluster on Die (COD) which splits the chip in two. This requires specific NUMA optimization, in that memory traffic to the other half of the chip can see a 3.7x decrease in bandwidth

Configuration Summary

Systems Under Test:

SPARC T7-1
1 x SPARC M7 processor (4.13 GHz)
512 GB memory (16 x 32 GB dimms)
Oracle Solaris 11.3 (11.3.1.5.0)
Java HotSpot 64-Bit Server VM, version 1.8.0_66
 
SPARC T5-2
2 x SPARC T5 processors (3.6 GHz)
512 GB memory (32 x 16 GB dimms)
Oracle Solaris 11.2
Java HotSpot 64-Bit Server VM, version 1.8.0_66
 

Benchmark Description

The benchmark description, as found at the SPEC website.

The SPECjbb2015 benchmark has been developed from the ground up to measure performance based on the latest Java application features. It is relevant to all audiences who are interested in Java server performance, including JVM vendors, hardware developers, Java application developers, researchers and members of the academic community.

Features include:

  • A usage model based on a world-wide supermarket company with an IT infrastructure that handles a mix of point-of-sale requests, online purchases and data-mining operations.
  • Both a pure throughput metric and a metric that measures critical throughput under service level agreements (SLAs) specifying response times ranging from 10ms to 100ms.
  • Support for multiple run configurations, enabling users to analyze and overcome bottlenecks at multiple layers of the system stack, including hardware, OS, JVM and application layers.
  • Exercising new Java 7 features and other important performance elements, including the latest data formats (XML), communication using compression, and messaging with security.
  • Support for virtualization and cloud environments.

Key Points and Best Practices

  • For the SPARC T5-2 server results, processor sets were use to isolate the different JVMs used during the test.

See Also

Disclosure Statement

SPEC and the benchmark name SPECjbb are registered trademarks of Standard Performance Evaluation Corporation (SPEC). Results from http://www.spec.org as of 11/30/2015.
SPARC T7-1 120,603 SPECjbb2015-MultiJVM max-jOPS, 60,280 SPECjbb2015-MultiJVM critical-jOPS; Cisco UCS C220 M4 97,551 SPECjbb2015-MultiJVM max-jOPS, 28,318 SPECjbb2015-MultiJVM critical-jOPS; Dell PowerEdge R730 94,903 SPECjbb2015-MultiJVM max-jOPS, 29,033 SPECjbb2015-MultiJVM critical-jOPS; Cisco UCS C220 M4 92,463 SPECjbb2015-MultiJVM max-jOPS, 31,654 SPECjbb2015-MultiJVM critical-jOPS; Lenovo Flex System x240 M5 80,889 SPECjbb2015-MultiJVM max-jOPS, 43,654 SPECjbb2015-MultiJVM critical-jOPS; SPARC T5-2 80,889 SPECjbb2015-MultiJVM max-jOPS, 37,422 SPECjbb2015-MultiJVM critical-jOPS; Oracle Server X5-2L 76,773 SPECjbb2015-MultiJVM max-jOPS, 26,458 SPECjbb2015-MultiJVM critical-jOPS; Sun Server X4-2 52,482 SPECjbb2015-MultiJVM max-jOPS, 19,614 SPECjbb2015-MultiJVM critical-jOPS; HP ProLiant DL120 Gen9 47,334 SPECjbb2015-MultiJVM max-jOPS, 9,876 SPECjbb2015-MultiJVM critical-jOPS; IBM Power S812LC 44,883 SPECjbb2015-MultiJVM max-jOPS, 13,032 SPECjbb2015-MultiJVM critical-jOPS.

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.