It is interesting to note that the price of server hardware, as a percentage of overall TCO (total cost of ownership), has decreased steadily over the past four years. Server hardware acquisition cost as a percentage of total TCO has been dropping as technology has advanced, and is now approaching only 20% of total three-year TCO.
Moreover, processor price as a function of TCO is now relatively constant across all vendors, with the processor price itself comprising only a very small, insignificant percentage of the total solution cost. With CPU costs basically equal and overall hardware costs steadily declining, the major TCO differences between vendors boils down to overall server performance, other hardware capabilities, power consumption, and the strength and scalability of the management stacks, particularly at scale.
HP has a long history of providing data center hardware and software solutions. HP’s blade server lineup dates back to RLX, which patented and shipped the first blade server in 2001. After HP acquired RLX in 2005, the company leveraged RLX blade technology to produce HP’s ProLiant blade server lineup. From a management perspective, the HP stack is comprised of a set of agent-based technologies either acquired from other sources or developed internally, all of which have been more or less integrated.
Cisco is a new entrant to the blade server market, first introducing the UCS blade server line in March 2009, although the company has had a dominant position in the data center networking market for years. While the UCS servers have been on the market only about a year, Cisco has already caused significant market disruption by introducing innovative blade server technologies that provide high performance with the advantage of organically developed management capabilities that are, by and large, built-in to the UCS hardware and firmware with a single interface.
CPU and Memory
As one might expect, CPU comparisons between the two vendors are virtually identical. Both vendors utilize the same Intel Xeon 5500 and 5600 series processors, both offer two-socket blades, and both currently offer (Cisco) or plan to offer (HP) four-socket Intel Nehalem EX-based blades. Both vendors use the same type of DDR3 DIMM memory, supporting 2GB, 4GB and 8GB modules.
There is one significant difference between Cisco and HP in terms of memory architecture, however.
HP’s conventional memory scheme supports either 12 or 18 DIMM slots, which translates to a maximum of 192GB of RAM per blade. Cisco’s “Extended Memory” architecture, developed in partnership with Intel, supports up to 48 DIMMs for a maximum of 384GB per blade. In addition to providing large amounts of memory for RAM-intensive applications such as VMs, Cisco’s memory architecture also enables the use of faster 1066 MHz chips across all 48 DIMMs, whereas HP servers currently drop to 800 MHz due to memory architecture.
Networking
When comparing blade vendors from a networking perspective, it is important to consider the overall supporting hardware required. The days of discrete networks for management, compute, and storage are fading fast thanks to the emergence of Fibre Channel over Ethernet (FCoE), which allows the convergence of all network protocols across a single medium.
HP’s blade server architecture, while making some steps towards convergence, is still firmly steeped in the old world. HP requires one or two separate network interface cards (NICs) per blade (one Ethernet and one Fibre Channel), a minimum of 4 “in-chassis” switches, plus associated cabling and downstream switches, to support each blade’s connectivity requirements. In addition to a large amount of supporting hardware and switches, this architecture requires a large quantity of network cables-up to ten cables per chassis (four Ethernet, four Fibre Channel, and two management ports per chassis) to reach a network aggregation density similar to Cisco UCS. HP’s new converged networking offering is delivered and managed “old world,” without true simultaneous convergence of network protocols with server, network and storage management.
Cisco brings converged Ethernet and FCoE directly to the blade chassis-a key advantage. This requires only four cables per chassis to support the same number of blades as HP while eliminating a large percentage of the downstream switching and physical cabling required. Converged fabrics carry the traffic for all networks, including compute, storage and management, through a 10 Gbps Ethernet connection.
Another major advance with Cisco UCS blades is their use of virtual interface card (VIC) architecture, allowing dynamic I/O configuration based on policies defined in UCS Manager. This allows the creation of up to 128 PCIe devices, configured as a vNIC (virtual NIC) or vHBA (virtual HBA). This is a significant concept, as it completely abstracts network configuration from the hardware, allowing it to travel with the Service Profile-something that virtually no vendor other than Cisco can do. This substantially reduces TCO and yields added benefits from increased network throughput and reliability, and major decrease in required management intervention when migrating VMs to other hosts; not to mention power savings from the reduced number of network components required.
UCS Manager is not only embedded in the UCS 6100 Series Fabric Interconnects (FIs), but also fully enabled in each blade with a dynamic dedicated private bandwidth for “Out Of Band” management connectivity, totally separate from production network traffic. The FIs scale across multiple blade server chassis in a single domain (potentially up to 40 chassis per FI). This eliminates need for expensive switches in every chassis, duplicate management servers, and database servers-a major savings compared with domain proliferation that can happen with HP at scale. UCS Manager therefore can manage up to 320 servers in a single domain structure regardless and independent of the total VM population.
Power Requirements
From a power perspective, there are a number of items to consider when comparing blade vendors. In addition to the raw amount of power consumed by servers and other supporting hardware (such as networking as outlined above), also consider the extra power required by management servers and consoles. This is particularly interesting at scale, where power required by management and supporting components must also scale. Comparing HP and Cisco, Cisco again has an advantage in this area from a number of perspectives. First, UCS requires only one management server, compared with up to four required by HP. Also, consider that each management server also requires an associated database server, driving power requirements even higher, not to mention the management burden required to maintain the servers and extra licensing costs.
From an architectural and management perspective, Cisco UCS compares very favorably against HP.
UCS Manager can power up a blade and provision it from bare-metal state within a few minutes using templates to define firmware on service profiles. This allows spare servers to be consolidated down to a fraction of the quantity required by other vendors-saving a lot of power while reducing overall hardware requirements and additional capital cost.
Blade Management Capabilites Hardware is important when comparing blade server vendors, but management overhead plays an even larger role in TCO than the hardware. Incumbent blade server vendors, HP included, base their blade server management approach on existing management technologies, which are comprised of a number of independently developed management technologies, integrated and extended to support new technologies like blade servers and controlled by a top-down orchestration engine. The result is an unfortunate mish-mash of somewhat integrated components that typically require multiple agents and consoles or applets, not to mention a lot of training in order for staff to understand all of the moving parts. In addition, incumbent vendors also tend to price their management components separately, further adding to the solution TCO.
Strong partnerships with Intel, EMC and VMware to build management components directly into Cisco began with a blank slate, taking advantage of their deep hardware management heritage and blade, chassis and networking hardware. Advantages of this approach include a single management agent on each blade, coupled with a true single management console. HP, along with most other blade vendors, requires a dedicated management blade for each chassis and multiple software management agents on each blade. This also requires multiple consoles or a single console view with multiple applets to access the various consoles.
Cisco’s architecture does not consume a blade for management, allowing one extra production server blade per 16 blades. They use a single, lean UCS agent, increasing overall blade capacity. Cisco includes the UCS management stack in the base price of its solutions with the exception of BMC BladeLogic, which is an OEM that provides bare-metal provisioning for Cisco and other blade vendor hardware.
Cisco integrates UCS Manager with many other management software companies’ products as well- not just BladeLogic. A key advantage of the Cisco architecture is that logic is pushed to the level of the action-when a blade is defined, it is given a set of policies that enable it to make intelligent, policy based management decisions without requiring the latency and overhead of an orchestrator.
Digging into HP’s blade management stack, one discovers a bewildering “alphabet soup” of components. HP requires the Virtual Connect Enterprise Manager (VCEM), Insight Control Environment (ICE) bundle, HP Systems Insight Manager (SIM), HP Integrated Lights-out (iLO) Web interface, HPOn-board Administrator, and HP Insight Dynamics VSE Suite, etc., which is comprised of up to 11 agents and up to four different management consoles. This requires a staggering amount of memory and processor overhead on every server, not to mention the administrative overhead required to maintain all of the management components, as the numerous interdependencies and complexities of the HP schema can cause failures if firmware versions fall out of synchronization.
Cisco UCS, in contrast to HP, requires a single Cisco UCS Manager console and a single UCS management agent per blade. Cisco also designed UCS Manager to work with more than just UCS blades. Cisco is adding support for Cisco rack mount servers to UCS Manager, which can also manage UCS network switches and Fabric Interconnect devices. When you consider the amount of additional installation, staff training and management overhead required for the HP blade solution, it is easy to see how Cisco comes out ahead from this perspective.
The one blade management technology that Cisco did not build in-house was bare-metal provisioning, which they chose to provide via tight integration with BMC BladeLogic, a highly regarded, agnostic, multi-vendor blade provisioning solution. Cisco fully integrated UCS with BladeLogic using open interfaces, performing all management tasks from the same UCS Manager console. In addition to simplifying provisioning management, BladeLogic integration also allows UCS to manage non-Cisco blades. Cisco UCS Manager is also fully integrated with other management software suppliers as well, such as CA, Symantec/Altiris, Microsoft, EMC , VMware and even HP and IBM. Interesting enough, Cisco provides full access to developers and uses standard XML API, so custom tools can be integrated with UCS Manager via a Cisco API SDK.
Conclusion
Cisco has created an impressive “next generation” data center architecture with UCS. Combining a dramatically simplified, intelligent hardware solution that abstracts servers, network and storage from the applications that depend on it with a centralized, policy-based engine drastically reduces management complexity. From Enterprise Management Associates’ perspective, Cisco wins the comparison with HP handsdown, for a number of reasons:
- Cisco UCS blades use ~10% less power than HP under load, and ~3% less power than HP when idle, potentially saving tens of thousands of dollars per year in energy costs.
- The UCS architecture allows many server spares to be turned off when they are not used, saving100% more energy than HP, and UCS allows fewer total server spares to be used than other vendors,since blades can be quickly and automatically provisioned from a bare-metal state, regardless ofthe application. This allows many more production servers per square meter of data center spacethan comparable blade servers.
- Bringing converged networking (including FCoE) directly to the blade chassis is a stroke of genius.The combination of eliminating the requirement for up to eight network cables per chassis, coupled with embedding the UCS manager in the chassis, blade and networking hardware, drastically reduces the TCO, power requirements and management complexity of the UCS solution.
- The UCS Manager ties all of this together from a single console, requiring only a single lightweight agent on each blade, compared with up to 11 agents and four consoles in the HP solution. Power savings and extra processing capacity derived from the elimination of management blades, extra management consoles and other supporting servers is significant.
- UCS hardware, which includes the ability to self-manage based on policies set by UCS Manager, allows management to be moved to the element level. This increases data center adaptability while decreasing management overhead and latency.
Organizations that are concerned with reducing environmental impact through decreased energy consumption and saving money on energy costs, reducing management overhead and increasing overall service levels, should take a hard look at Cisco’s Unified Computing System.
While UCS represents evolutionary advances in hardware technologies, combining UCS hardware with the innovative UCS management stack shows great promise that could very well make a revolutionary impact on the data centers of today and in the future.
http://i.zdnet.com/whitepapers/Cisco_ema_data_centermgmt0610.pdf
