Original URL: http://www.theregister.co.uk/2010/03/16/intel_westmere_ep_pitch/

Intel wants vintage x64 servers on rubbish heap

Time to buy Westmere, says Westmere maker

By Timothy Prickett Morgan

Posted in HPC, 16th March 2010 15:49 GMT

With the economy on the mend - at least by gross measures that may not mean a hill of beans to people on the street - Intel thinks it is putting its "Westmere-EP" Xeon 5600 processors into the field at precisely the right time. Not only are people more willing to spend money than they were a year ago, but their aged servers are one year older and closer to death.

According to Boyd Davis, general manager of marketing for Intel's Data Center Group, the new six-core Xeon processors make an even better case for the removal of vintage single-core and dual-core x64 servers and the consolidation of their workloads onto bare metal or virtualized Xeon 5600 boxes than the "transformational" Nehalem-EP Xeon 5500s did a year ago.

Last March, the Xeon 5500s, thanks largely to the factor of three expansion of memory bandwidth that came from the ditching of the old Xeon frontside bus architecture for the new QuickPath Interconnect, represented the largest performance boost in the history of Intel server chips.

The quad-core Xeon 5500s outperformed their quad-core predecessors, the "Harpertown" Xeon 5400s, by anywhere from 70 to 125 per cent, depending on the workload, with memory-sensitive workloads seeing the most gains. Such performance increases, coupled with aging x64 machinery, is what compelled customers to start spending again on x64 servers in the second half of 2009.

The six-core "Istanbul" Opteron 2400 and 8400 processors did their part as well in the x64 server recovery by delivering a roughly 50 per cent performance improvement in existing quad-core "Shanghai" Opteron machines. This time around, Intel has the existing Nehalem-EP machines and is doing a drop-in upgrade that offers roughly the same performance bump and also pushes the core count up to six per socket, and in a few weeks, it will be AMD that is rolling out a new chipset and a tweaked architecture, pushing per-socket core counts to a dozen with the "Magny-Cours" Opteron 6100s.

The argument that Intel was making a year ago, according to Boyd, was the Xeon 5500s offered around nine times the performance of old single-core Xeons sold in 2005. So moving up to new Nehalem-EP boxes could pay for itself in about eight months when you take into account the reduced number of software licenses and lower power, cooling and maintenance expenses for the newer systems. (That scaling factor was based on the SPECjbb_2005 Java benchmark test, which is the most friendly to such a consolidation measure thanks to the Java virtual machine and its virtualized runtime).

Now, with the Westmere-EP chips, using the same Java consolidation methodology, Boyd says that customers can get 15 times the oomph per machine. Which means a few different things. First, they could replace the old boxes, rack for rack, with machines based on Xeon 5600s and get 15 times the performance and still cut their energy bills by around eight per cent. Or, they could compress their server racks by a factor of 15 and still have the same amount of raw computing, cut their annual power bills by 95 per cent, and get a return on investment in around five months.

By Intel's math, you can take 50 of those single-core, two-socket vintage Xeon boxes and replace them with three two-socket Xeon 5600 servers, and save $5,092 per month on software support, $1,838 per month on electric bills, and $3,125 per month on maintenance fees, assuming the box is running Red Hat Enterprise Linux alone. That's $10,055 per month. (You can use Intel's Server Refresh Estimator to run your own scenarios and see what the savings would be from other gear and using other workloads.)

Intel colored glasses

The scenario that Intel painted above is a particularly rosy one, given that Java performance gains over the Xeon 5500s are highest for the Xeon 5600s, an increase of 46 per cent according to Intel's tests. Also painting some more pink on the lenses in Intel's benchmark comparisons is the fact that it pitted a six-core 130 watt Xeon X5680 part running at 3.33 GHz against a four-core Xeon 5570 running at 2.93 GHz to come up with its benchmark ratings.

That's not exactly a kosher comparison, since very few server buyers will go anywhere near the 130 watt parts and will be happier with the 95 watt Xeon 5600s, which top out at the same 2.93 GHz with the Xeon X5670. If performance scales roughly with clock speed, that clock speed differential (3.33 GHz versus 2.93 GHz) accounts for about 14 per cent of the performance increase of the move from Xeon 5500s to Xeon 5600s and also helps with consolidation ratios from older machines, too.

You have to read the fine detail with all chip makers.

On what Intel calls mainstream enterprise applications, SAP ERP performance on the top-end Xeon 5680 was only up 27 per cent compared to the Xeon X5570, and raw integer performance is up 40 per cent, virtualization VM guest count is up 42 per cent based on the VMark benchmark, and energy efficiency (as gauged by the SPECpower_ssj2008 benchmark) was up 42 per cent.

For the propellerheads, memory bandwidth is up by 20 per cent, floating point math is up by 25 per cent, and on many HPC and financial services applications, performance improvements on a two-socket server using the 3.33 GHz Xeon 5680 are more than 60 per cent higher than on a machine using two 2.93 GHz Xeon 5570s. (These tests included the Linpack Fortran benchmark commonly used on supercomputers as well as Black Sholes risk calculations commonly used by financial institutions.)

In a like-for-like comparison, Intel pitted the X5570 against the X5670, both running at 2.93 GHz and both rated at 95 watts, and the latter chip delivers 40 per cent more performance per watt on the SPECpower_ssj2008 benchmark. For some reason, Intel used the SPEC_int_rate2006 test to do a energy efficiency comparison, saying that a newer low-voltage six-core Xeon L5640 running at 2.26 GHz and rated at 60 watts could deliver the same performance as a 2.93 GHz Xeon X5570 while consuming 30 per cent less electricity.

As El Reg has previously reported, the Westmere-EP chips include instructions for encrypting and decrypting data using the Advanced Encryption Standard (AES) algorithm. And Boyd says the significance of this is huge. Using a Web-based banking application, Intel took a Xeon 5500 server from last year and found it could support around 13,000 users.

Turning on SSL encryption, which uses the AES algorithm, reduces the number of banking users the system can support to 10,500 - a 19.2 per cent performance hit for turning on security. But moving to a two-socket box (again, using that hot 3.33 GHz Xeon X5680 instead of the 2.93 GHz X5670) allows the banking app to have SSL encryption turned on and support 16,000 users.

That is much more dramatic than admitting that the extra performance embodied in the two extra cores in the Westmere-EPs will be burned up supporting SSL. (Yes, I know the AES is actually in all six cores). But what is true is that now Intel and its server partners can make a compelling case to that additional cryptographic processors are not necessary in x64 servers.

While all of this helps the Intel Westmere-EP sales pitch, nothing perhaps helps more than the fact that according to the analysts at Gartner, about a million servers that would have been replaced in 2009 were not, and another bunch of servers will need to be retired this year, no matter what. By Intel's estimate, about a third of the x64 installed base is still comprised of single-core x86 and x64 boxes and that a whopping 80 per cent of the installed base, when you toss in dual-core machines and even some older quad-core machines, are up for a refresh this year.

That doesn't mean those tens of millions of machines will be replaced, but it does mean that Intel and its server partners are going to make the strongest case possible that they ought to be. ®