Preview:Intel Harpertown Quad Core CPU

This month we got an opportunity to get our hands on the latest Intel's 45 nm
based Xeon processor codenamed Harpertown. We received a complete rack mountable
server with two 3.0 GHz quad core Harpertown processors, 16 GB RAM, and two 200
GB Sata hard-disk connected through a PCI-X RAID controller. We ran a set of
tests to see the performance of the server and got some really amazing results.
But before we talk about the tests, results, and performance of the server,
let's first understand what is this new processor 'Harpertown' all about.

What is Harpertown?

With the launch of the 45 nm Penryn processors, Intel has announced their
Penryn based Xeon (5400 series) models codenamed Harpertown, with higher front
end bus speed, 1600 MHz. This series of processors will be available from 2.6 to
3.2 GHz range. But the 2.6 GHz one will have a lesser bus speed which is 1333
MHz. Intel is even planning to launch a dual core version of the CPU with a
similar architecture which will be codenamed as Wolfdale and will be available
with frequency ranging from 1.89 to 3.4 GHz. In simple terms Harpertown is a
better, faster, and more energy efficient quad core processor if compared
against Clovertown (the current quadcore based Xeon CPUs). The key difference is
in the L2 cache, bus speed, and power consumption. For a detailed comparison
have a look at the table. Harpertown is supposed to replace Clovertown in the
near future.

Another highlight of Harpertown is its new instruction set. Harpertown comes
with Intel Streaming SIMD Extensions 4 (SSE4) instructions. This instruction is
today the largest unique instruction set addition since the original SSE
instruction set architecture was first used.

Setup and tests

To test the Harpertown in our test labs we ran benchmarks such as Linpack,
SunGard, and Netbench and our own developed BZip2 benchmark. We ran these tests
on multiple OSs such as CentOS 5 and Windows Longhorn Server RC0. The setup was
pretty much simple. We took the server as it is and only changed the RAM in some
cases. The only configuration made was a RAID 0 array of the two hard disks to
get the best possible read-write performance from the 5400 series server.

Test 1: Running Longhorn

We installed Longhorn on this server as this one would be the latest OS when
Harpertown would be released in the market. And most of us will be using
Harpertown Servers with Windows Longhorn pre-installed into it. So testing
compatibility and performance of Longhorn server was a must for us. While
installing Longhorn we really got surprised as the complete installation of
Longhorn took less than 10 minutes which generally takes around half an hour on
a core 2 duo extreme machine. We also gave a try to Longhorn's new hypervisor on
Harpertown and it worked marvelously with full virtualization support. We were
able to run seven dedicated guest systems on the hypervisor.

Test 2: Linpack

A server testing without Linpack doesn't make any sense. But this time
instead of running Linpack on a stripped down installation of OS with minimum
service running, we decided to run it on top of a full blown Longhorn
installation itself. We did this to see how much of actual performance a user
will be achieving while using Harpertown on a production environment with an OS
such as Longhorn.

Sungard risk analysis application took 203.12 seconds to complete the test job with the 8 core Harpertown server

With a bit of tweaking in the Linpack input file we were able to achieve
around 65 Giga Flops in no time. This result is till now the best performance
result we have got out of any server. The closest performer to this score was
the Connoisseur Skyrunner which was a dual Clovertown 3.2 GHz based server which
gave us around 53 Giga flops of floating point operations.

And as we ran Linpack on a full blown OS without stripping down any services,
we can easily say that after fine-tuning the OS we can easily achieve up to 70
to 80 Giga Flops sustained performance with Harpertown.

Bzip2

We also ran our self-developed Bzip2 benchmark which zips a huge number of
files simultaneously and stresses the processor and the hard-disk. The
performance which we got was not as good as we achieved in our Clovertown based
Connoisseur Skyrunner server. But the reason was apparent. The Bzip2 is a
benchmark which stresses both the processor and the hard-disk. The Harpertown
server, which we received, had just two disks whereas the Skyrunner had four in
a RAID 0 array which gave it an edge over this Harpertown server.

SunGard

SunGard is a financial risk analysis application which is also used for
stressing the processors. The application even lets you select the number of
cores which we want to strain. We ran this test on the Harpertown and ran it on
various cores to see the performance increase is linear or not when we increase
the cores. The result was not exactly linear. We saw a sudden drop between two
and four cores but the drop between six and eight was the least. For more
details see the SunGard graph.

The 'performance increase' with the increase in the number of cores was not linear

We also decided to see the per-core performance difference in a core 2 duo
extreme and Harpertown machine. To test this we took a Core 2 Extreme 3.0
machine with 2 GB of RAM and ran SunGard on it. So, essentially SunGard ran with
two processing threads. Then we took the Harpertown server and reduced its RAM
from 16 to 2 GB to make sure that the benchmark doesn't outperform just because
of the extra RAM it has. Then we ran SunGard on it too. We got around 15% of
performance benefit in the Harpertown 2 Core test which is a good result.

Bottomline: With the best performance among all the quad core CPUs, we have
reviewed, this processor is the choice of the future.

Harpertown gave 15% performance benefit in a machine running two active cores than a core 2 duo extreme machine