White Papers

ManualsBrandsDell ManualsConverged InfrastructureHigh Performance Computing Solution Resources

Eeny Meeny Miney Mo, should I go with 20 cores?

Nishanth Dandapanthula, June 2014

Intel’s Xeon E5-2600 v2 product family processors (architecture code named Ivy Bridge) have been

available in the server market for a few months now. Ivy bridge processor based systems provide better

performance when compared to previous generation processor families such as Sandy Bridge (Xeon E5-

2600) and Westmere (Xeon X5600). This can be attributed to several factors, such as increased core

counts because of the 22nm process technology, higher clock rates, increased system memory speeds and

larger last level cache. This performance improvement is shown in several studies mentioned in 1, 2 and 3.

So, once the decision to move to a new platform or new processor technology has been made, what next?

How should these new systems be configured? There are so many choices for the processor itself –

different options with different core counts, processor frequency, TDP and, of course, price. Which

processor model is optimal for a specific workload? This blog provides quantitative data and analysis to

help answer this question by comparing the performance and power profile of different processor models

across a variety of HPC applications.

Once the decision has been made to choose a particular processor SKU, there is another important

decision which needs to be made. This is in regard to the choice between single rank and dual rank

memory modules. As mentioned in studies 4 and 5, the choice and configuration of memory modules

impacts the performance and power consumption statistics of an environment. This blog also describes a

performance comparison done between single and dual rank memory modules for several HPC

applications. These studies were done in the Dell engineering lab in December 2013 and results are actual

measured results.

Table 1 describes the configuration of the HPC test cluster and the benchmarks used for this analysis. Six

different types of Ivy Bridge processors were studied along with two types of memory modules, single

ranked and dual ranked.

BIOS options were set to reflect Maximum Performance settings, details are in Table 1. A four node

cluster was interconnected with InfiniBand. All of the results shown in this blog were obtained by fully

subscribing the four servers, i.e., all cores on all servers were in use.

Table 1 Test bed configuration and Benchmarks

Servers

4 x Dell PowerEdge C6220 II sleds in one chassis

Processors per server

2 x E5-2643 v2 @ 3.5 GHz 6c 130 W

2 x E5-2667 v2 @ 3.3 GHz 8c 130 W

2 x E5-2680 v2 @ 2.80 GHz 10c 115 W

2 x E5-2670 v2 @ 2.5 GHz 10c 115W

2 x E5-2690 v2 @ 3.00 GHz 10c 130 W

2 x E5-2697-v2 @ 2.7 GHz 12c 130W

Memory per server

1 DPC 8 x 16 GB 1866 MT/s Dual Rank

Summary of content (7 pages)

PAGE 1
Eeny Meeny Miney Mo, should I go with 20 cores? Nishanth Dandapanthula, June 2014 Intel’s Xeon E5-2600 v2 product family processors (architecture code named Ivy Bridge) have been available in the server market for a few months now. Ivy bridge processor based systems provide better performance when compared to previous generation processor families such as Sandy Bridge (Xeon E52600) and Westmere (Xeon X5600).
PAGE 2
1 DPC 8 x 8GB 1866 MT/s Single Rank Mellanox ConnectX3 FDR adapter Mellanox SX 6036 FDR Switch Interconnect Turbo Enabled BIOS Options C States Disabled Hyper Threading Disabled Node Interleaving Disabled Power Management Max Performance Mellanox OFED 1.5.3-3 OFED version RHEL 6.4 (kernel 2.6.32-358.el6.x86_64) OS 2.1.2 BIOS version 2.25 BMC version HPL 2.0 N set to 85 % of total memory Benchmarks available NB = 168 NAS PB 3.3.1 LU Class D Stream for Memory Bandwidth N = 160000000 Open Foam 2.2.
PAGE 3
Figure 1 Performance across four nodes using multiple IVB processors and dual rank DIMMs Figure 1 compares the different processor models for each benchmark from a pure performance standpoint. The graph plots cluster-level relative performance for each benchmark using the Xeon E52643 v2 processor (3.5 GHz 6 core) as the baseline. Dual Rank DIMMs were used for this study. From the graph it can be seen that best performance is obtained with the 12 core E5-2697 v2 (2.
PAGE 4
The other aspect is processor frequency. Once the memory bandwidth requirements are met, faster processors will perform better. This is clearly illustrated in the 10c case with the 3.0 GHz processors performing better than the 2.8 GHz parts, which in turn perform better than the 2.5 GHz model. Thus in the LS DYNA scenario, the E5-2667 v2 performs the best, providing the right balance of memory bandwidth and processor speed.
PAGE 5
v2 10 core part is doing the best in terms of performance per watt but the E5-2667 v2 processor does the best in terms of pure performance. The difference in this case is less than 5 %. This could be because of lower wattage of the E5-2670 v2 processor (115 Watts). Figure 3 Performance drop when using single rank memory modules As mentioned previously, choice of memory modules can affect application performance. The study done in this blog shows that Dual rank memory modules give the best performance.
PAGE 6
Figure 4 Performance per core across four nodes for multiple processors using dual rank memory modules Figure 4 shows the application performance in terms of performance per core. The 3.5 GHz E5-2643 v2 processor has the highest clock speed among all the processors used and gives the best performance per core.
PAGE 7
1. http://en.community.dell.com/techcenter/high-performancecomputing/b/general_hpc/archive/2013/11/07/performance-comparison-of-intel-xeon-e5-2600and-e5-2600-v2-series-processors.aspx 2. http://en.community.dell.com/techcenter/high-performancecomputing/b/general_hpc/archive/2012/03/06/hpc-performance-on-the-12th-generation-12gpoweredge-servers.aspx 3. ftp://ftp.dell.com/Manuals/Common/poweredge-r420_White%20Papers96_en-us.pdf 4. http://www.dell.