Tuesday, June 09, 2009
Thursday, June 04, 2009
Intel Xeon 5500 Series Processors for Servers and Workstations
There are a few key things you have to know about the 5500 series before you design your next server or workstation.
First you need to know that they are an order of magnitude better than the 5400 series, offering over 100% better performance with signicantly lower power consumption. Why its faster, is not the purpose of this article, but that information can be found here. Helping you understand the models, and how to architect a system around them is the purpose of this article.
There are a number of models, that fall into three categories.
1) The models that start with 550 are entry level models with 4MB L3 Cache and use DDR3 800 memory. They all have four cores except for the 5502, which has two.
2) The Models from the 5520 to the 5540 have 8MB L3 cache, use DDR3 800 or 1066 memory, and have the added benefit of Turbo Boost and Hyperthreading.
3) The top models starintg with the 5540 also can use DDR3 1333 memory, and have significantly more potential benefit from Turbo Boost.
Before we discuss the memory architecture, a quick summary of the other features.
Hyperthreading, as you might recall from prior processors, is the abilty to run two threads per core, to better saturate the processor with work. With two Quad Core 5500 series processors and hyperthreading, you get 16 threads. This is really helpful for Virtualization.
Turbo Boost is a new concept. Turbo Boost allows the processors to overclock automatically if doing so will not cause the processor to exceed its rated power consumption and thermal footprint. For example if an application that runs on one or two cores is running, the processor will power down the unused cores, and boost the performance of the remaining cores. The bottom line, is the better the thermal design of the server and the better cooled the environment, the more effective Turbo Boost will be.
Memory architecture is very important when planning systems around the 5500 series. Here is how it works.
Each processor has three paths to memory, so it peaks its performance if you purchase 3 DIMMS per processor. In a dual processor system you should consider 6, 12, and 24GB to be top performance for the dollar, using 1GB, 2GB, or 4GB modules.
Another key concept is that each processor can have up to three banks of memory, but there is a substancial performance penalty if you use banks two and three. If you use one bank of 3 modules per processor, and if you use one of the top category processors, then can benefit from the purchase of DDR3 1333MHz memory.
But, if you place even one more module in the system, all of the memory drops down to 1066. If you place any memory in a the third bank, your memory will all operate at 800MHz.
Because expanding beyong the first bank (3 per processor, 6 in total DIMMs) causes a serious reduction in speed, I recommend purchasing the largest modules that make finanfial sense (2 or 4GB), and populate as many sockets as you need in the first bank. Go to 8GB modules if you need lots of memory, but they are currently very expensive.
If performance really matters then populate all six sockets in a 2 processor system. If performance is not an issue yet, its better to buy fewer higher capacity modules now so you can upgrade later without reducing performance.
Most servers don't require the top processors. For those servers, using more banks of memory will not necessarily entail a loss of perfomance. The 550X processors only support 800MHz memory anyway, so they suffer no penalty from using all the memory banks. The next level support up to 1066MHz, so they suffer no penalty from using 2 banks or 6 DIMMs per processor.
In all cases you maximize your data path by using exactly three DIMMS per bank per processor.
Finally when a motherboard has only one processor, make sure you do not populate the second processor's memory banks, as they will not be available. Conversely if you have two processors, make sure you distribute the memory evenly between them.
Best of all let Ashlin help you design your servers based on your specific requirements, both now, and in the future.
First you need to know that they are an order of magnitude better than the 5400 series, offering over 100% better performance with signicantly lower power consumption. Why its faster, is not the purpose of this article, but that information can be found here. Helping you understand the models, and how to architect a system around them is the purpose of this article.
There are a number of models, that fall into three categories.
1) The models that start with 550 are entry level models with 4MB L3 Cache and use DDR3 800 memory. They all have four cores except for the 5502, which has two.
2) The Models from the 5520 to the 5540 have 8MB L3 cache, use DDR3 800 or 1066 memory, and have the added benefit of Turbo Boost and Hyperthreading.
3) The top models starintg with the 5540 also can use DDR3 1333 memory, and have significantly more potential benefit from Turbo Boost.
Before we discuss the memory architecture, a quick summary of the other features.
Hyperthreading, as you might recall from prior processors, is the abilty to run two threads per core, to better saturate the processor with work. With two Quad Core 5500 series processors and hyperthreading, you get 16 threads. This is really helpful for Virtualization.
Turbo Boost is a new concept. Turbo Boost allows the processors to overclock automatically if doing so will not cause the processor to exceed its rated power consumption and thermal footprint. For example if an application that runs on one or two cores is running, the processor will power down the unused cores, and boost the performance of the remaining cores. The bottom line, is the better the thermal design of the server and the better cooled the environment, the more effective Turbo Boost will be.
Memory architecture is very important when planning systems around the 5500 series. Here is how it works.
Each processor has three paths to memory, so it peaks its performance if you purchase 3 DIMMS per processor. In a dual processor system you should consider 6, 12, and 24GB to be top performance for the dollar, using 1GB, 2GB, or 4GB modules.
Another key concept is that each processor can have up to three banks of memory, but there is a substancial performance penalty if you use banks two and three. If you use one bank of 3 modules per processor, and if you use one of the top category processors, then can benefit from the purchase of DDR3 1333MHz memory.
But, if you place even one more module in the system, all of the memory drops down to 1066. If you place any memory in a the third bank, your memory will all operate at 800MHz.
Because expanding beyong the first bank (3 per processor, 6 in total DIMMs) causes a serious reduction in speed, I recommend purchasing the largest modules that make finanfial sense (2 or 4GB), and populate as many sockets as you need in the first bank. Go to 8GB modules if you need lots of memory, but they are currently very expensive.
If performance really matters then populate all six sockets in a 2 processor system. If performance is not an issue yet, its better to buy fewer higher capacity modules now so you can upgrade later without reducing performance.
Most servers don't require the top processors. For those servers, using more banks of memory will not necessarily entail a loss of perfomance. The 550X processors only support 800MHz memory anyway, so they suffer no penalty from using all the memory banks. The next level support up to 1066MHz, so they suffer no penalty from using 2 banks or 6 DIMMs per processor.
In all cases you maximize your data path by using exactly three DIMMS per bank per processor.
Finally when a motherboard has only one processor, make sure you do not populate the second processor's memory banks, as they will not be available. Conversely if you have two processors, make sure you distribute the memory evenly between them.
Best of all let Ashlin help you design your servers based on your specific requirements, both now, and in the future.
Subscribe to:
Posts (Atom)
