Leveraging AMD's dual-core CPUs and the latest in HyperTransport ASIC technology, Appro goes Xtreme with a new line of servers poised to shakeup high-end datacenters trying to cope with the demands of server consolidation 'and virtualization.

by Jack Fegreus
April 3, 2006

F

Nonetheless, too many distributed servers can easily trigger a number of cascading negative effects. Rack mount servers are relatively inexpensive to acquire and easy to configure; however a high-density server deployment can quickly add up to considerable costs in floor space, management, networking, power, and heat. While chip areas have fallen and computational power increased, power consumption and heat dissipation have become more problematic. Increasing the number of servers also increases variability and lack of consistency in the data stored on these systems, which can create serious problems for IT when attempting to develop services relying on data from different systems.

With business-organization gurus obsessed with the “customer experience," IT strategies are once again turning to server consolidation schemes. Accelerating that trend are the Total Cost of Ownership (TCO) models that point to the costs of operating a server as dwarfing the costs of acquiring a server. The Robert Frances Group pegs the acquisition cost of a server to be only 20% of its TCO. This puts the spotlight on hardware resource utilization as the key to cost containment. It also puts IT under closer scrutiny to improve resource efficiency.

At the same time, many CEOs, who must dealing with issues such as shrinking product life cycles and the pursuit of global markets, have begun to institute real-time sense-and-respond strategies for business operations. Those strategies elevate the premium on server performance and leave CIOs on the horns of a dilemma: To drive down costs, they must increase server utilization; but to meet the processing requirements of a sense-and-respond business strategy, they must increase real-time performance. Successful resolution of those conflicting issues requires a sophisticated hybrid approach to traditional scale-up and scale-out strategies and a new paradigm in server design.

The dual-core AMD 280 and 880 CPUs are clocked at 2.4GHz. Differences between the two are a direct result of the direct-connect HyperTransport microarchitecture utilized by AMD. HyperTransport technology replaces the 20-year old Northbridge/Southbridge shared-bus model with high-bandwidth, low-latency, flow-through tunnels that pass data at up to 12.8GB per second using two unidirectional point-to-point links. Connecting bridges to tunnels provides a mechanism to create switch and star HyperTransport topologies. For 2-way systems, AMD 280 CPUs provide for one HyperTransport inter-processor connection and for 4- or 8-way systems the 880 CPUs provide for four inter-processor connections.

For comparison we also set up a 4-way server that featured single-core CPUs, an HP ProLiant DL580 G3 server. Like the Appro XtremeServers. the HP ProLiant DL580 G3 server was configured with 8GB of DDR2 memory clocked at 400MHz. The HP ProLiant DL580 sported 3.33GHz Intel Xeon CPUs with Extended Memory 64 Technology (EM64T). EM64T CPUs allow the system to address up to 1TB of combined virtual and physical memory via a flat 64-bit virtual address space, 64-bit pointers, 64-bit general-purpose registers, and 64-bit integer support.

On each of the three servers, openBench Labs installed SUSE Linux 10. Unlike SUSE Linux Enterprise Serve (SLES), which has a guaranteed 5-year life cycle for support, SUSE Linux is released every six months with an extensive set of emerging-technology applications for both server and desktop environments. SUSE Linux is intended as a technical preview of what Linux production environments will likely be running in the near future.

To help our assessment of Appro’s XtremeServers, the most important new technology introduced in SUSE 10 is version 4.0 of GCC, which now designates the GNU Compiler Collection. Historically, the moniker "GCC" explicitly stood for the GNU C Compiler; however, the acronym now properly designates the complete distribution—from language front ends, to language-independent optimizers, through to support libraries of GNU programming languages, including C, C++, Objective-C, Objective-C++, Java, Fortran, and Ada.

The most extensive and important change may be the overhaul given to the language-independent optimizers. Every component of the SUSE Linux 10.0 distribution has been compiled and checked with GCC 4.0. SLES 10, which will be released in late spring of this year, will be built upon the latest version of the new GNU Compiler Collection and will target mission-critical use.

We started our assessment by compiling single-threaded 64-bit versions of our oblCPU v4.5 benchmark that were optimized for execution on their target platforms: either AMD64 or EM64T CPUs. In addition to using GNU C 4.0, we used the Intel C v9.0 compiler to generate code for optimal execution on the Xeon EM64T-based server.

Using the new GNU C 4.0 compiler, AMD64 Opteron CPU performance scaled along clock-speed lines set by an Intel Pentium III CPU. This was not the case for the superpipelined Xeon EM64T CPU. In the jump from Pentium III to Pentium 4 processor architecture, from which Xeon CPUs have evolved, Intel instituted a superpipelined architecture to exploit instruction level parallelism without involving any special coding by the programmer.

With GNU C 4.0, the 3.3 GHz Xeon EM64T CPU provided only 75% of the CPU performance of the 2.4GHz Opteron 880 CPU. Only by usng the Intel C 9.0 compiler were we able to properly exploit the Xeon’s superpipelined architecture and approach the performance of the AMD64 Opteron CPU.

To assess the scalability of SMP load performance, we launched multiple copies of our single-threaded oblCPU benchmark on each of the three servers. The results were quite dramatic.

In the cases of the 2-way XtremeServer and the HP Proliant DLT580 G3, the difference in the net SMP load performance between the two servers paralleled the difference in performance between single instances of oblCPU. As a result, the two-way XtremeServer with dual-core AMD 280 CPUs supported a greater net SMP processing load compared to the 4-way HP Proliant DL580 G3 server with four single-core Intel Xeon EM64T CPUs and hyperthreading.

In these openBench Labs tests, job scheduler effects were the same on each system as all ran the same build of SUSE Linux 10 on the 2.6.13-15 Linux kernel.

When we normalized the net SMP load performance, the SMP load profile for the 2-way Appro XtremeServer and the 4-way HP Proliant DL580 G3 emerged as being identical. While hyperthreading makes the 4-way HP Proliant DL580 G3 with Xeon EM64T CPUs appear to the Linux operating system to be an 8-way server, our SMP load benchmark results demonstrated that its scaling profile was nothing like that of the 4-way Appro XtremeServer, for which dual-core CPUs provide true 8-way SMP load scalability.

To realize all of that potential bandwidth, however, it will be necessary to access multiple banks of memory under the control of multiple CPUs. For application programmers—not to mention all of the existing application programs—that smart microarchitecture has all the potential to become a complex programming nightmare.

To avoid just such a disaster, AMD64 technology makes the intricacies of NUMA transparent at the operating system level. At boot time, the AMD64 multiprocessor system BIOS calls on each processor to report its local memory. The system BIOS then maps each of the local memory pools into a single globally-addressable physical memory space. What's more, the AMD64 processors automatically maintain cache coherency across that global address space. As a result, programs access any data belonging to any processor with normal memory operations to the global address space. For programmers and system administrators, the only difference between NUMA and a traditional SMP system is the speed.

With that level of performance and the ability to support up to 64GB (2-way) and 128GB (4-way) of memory. the Appro XtremeServers are equally well-suited for memory intensive business tasks such as OLAP business intelligence and data mining, as they are for performing computational physics and engineering simulations.

 For PCIe-switched I/O, the 4-way Appro Xtreme server features nVidia’s nForce Professional 2200 and 2050 ASICs for PCIe. The nVidia nForce Professional 2200 also acts as a bridge for internal I/O connections including SATA, for which this server provides six hot-swap bays. We populated those bays with Western Digital 10K Raptor drives—one of which was reserved for SUSE Linux 10 Professional and the remaining five were combined as a high-throughput software RAID 0 device.

Running our benchmark for disk throughput, oblDisk v3.0, we pegged the SATA array as capable of sustaining extraordinary levels of continuous throughput, which averaged 155MB per second on writes and 252MB per second on reads. We also configured the four hot-swap bays on the HP ProLiant DL580 with 15K Ultra320 SCSI drives. which were connected to an HP StorageWorks 6i SmartArray controller. Similar to the Appro XtremeServer, we configured one drive for the OS and the remaining drives were put into a hardware-based RAID 0 array. As with previous CPU performance and memory bandwidth benchmarks the 4-way Appro XtremeServer held throughput advantages on reads and writes that were on the order of 3.5- and 1.5-to-1.