Date: August 13, 2005
The secret to sizing a server with Advanced Power Virtualization (APV) is to size CPU.s for the average workload. This is in contrast to a traditional server (or LPAR), which is sized for peak workloads.
Advanced Power Virtualization allows a physical server to run multiple partitions, sharing CPUs, boot disks, network and SAN adapters. Sharing uses the hardware more efficiently. Consequently less hardware is needed for the same workload.
The attached example illustrates how I would size an APV server to support the equivalent workload of 10 p510.s. I.ve quantified some of the benefits, which in this example reduced hardware costs in excess of 20%, and reduced power, cooling and rack space by over 80%. All considered, APV more than pays for itself.
Standard disclaimer: each situation is different, your results may vary.
Assume there are 10 p510.s, each having 1x1.65GHz CPU and 1 GB memory. Each p510 is sized for the peak loads, with an expected average CPU utilization of 20% (Gartner.s industry average utilization rate for Unix servers). Each p510 has dual Ethernet and dual boot disks.
CPU Sizing and Model Selection: The CPU.s for APV are sized for the average combined workload of all the partitions, plus a shared free pool of CPU resources for peak processing. A small VIO partition is also required to support shared boot disks and Ethernet adapters.
The sizing metric is IBM's rPerf performance rating. The p510 has a rPerf rating of 5.24. The expected average utilization is 20%. So the aggregate average rPerf for 10 servers, and VIO server would be:
p510 rPerf: 10.48 (10 servers)*(5.24 rperf/server)*(20% util) VIO rPerf: 1.0 ------------------------------------------ Average rPerf = 11.48
Additional capacity is required for the shared CPU pool for peak processing. As a first pass, I size the free pool by specifying the overall server utilization. A reasonable number would be 60%. This results in a overall rPerf rating of 19 (11.48/60%). A good APV server would be the p550 which has an rPerf rating of 19.66 (4x1.65GHz).
Memory: The total memory requirement is 10 GB (10 servers * 1 GB/server). Rounding up to 12 GB provides the additional memory for partitioning overhead (700MB), and matches available memory hardware increments.
Boot disks: The client partitions will boot off shared .virtual disks. located on the VIO server. A total of 8-72 GB drives will be sufficient. Two disks (mirrored) will be allocated to the VIO server operating system, and the remaining 6 disks (420GB) will support the AIX requirements of the 10 partitions with 100% spare capacity.
This assumes 20 GB disk per AIX partition and modest I/O rates. AIX will comfortably fit on 10 GB disk (OS + paging), or 20GB with redundancy. .Modest I/O. rates means no paging or heavy I/O as in a database.
Network: Client partitions will share 2 redundant GigaBit Ethernet adapters located on the VIO server.
|CPU's (per server)||1 x 1.65 GHz||4 x 1.65 GHz|
|Memory||1 GB||12 GB|
|Boot disk||2x72 GB||8x72 GB|
|Expected CPU Util.||20%||58%|
|P510 Alternative||P550 APV Alternative||APV Savings|
|Max power (watts)||6000||1100||82%|
|Max cooling (btu)||20460||2557||87%|
|Rack Space (height)||35||7||80%|
In this example, APV reduced hardware costs by $17,000 or 25%. Other tangible savings include power, cooling and rack space. Depending on charge back rates, this translates to savings of around $10,000 per year.
Another significant benefit is the software savings, which is often overlooked in the hardware purchase. Many software applications (Oracle, DB2, Websphere) are priced based on the number of CPU.s. In this example, APV would reduce this software cost by 60% because we.re using 4 versus 10 CPU.s. To provide an order of magnitude savings, assume the software costs $20,000 per CPU. If so, the savings would be $120,000 - more than double the cost of the hardware itself! In essence, the hardware is paid with the software savings.
Finally, APV flexibility reduces the sizing risk compared to standalone servers. CPU resources are .load leveled. dynamically. Undersized partitions will dynamically grab CPU cycles from the shared free pool while oversized partitions will donate unused cycles back to the free pool.
Memory resources can be redistributed between partitions. Memory granularity is in MB, and the reallocation is dynamic.
When using Advanced Power Virtualization, servers are sized based on the aggregate average workload, plus a smaller shared pool of CPU.s for peak processing. APV requires a paradigm shift in thinking, but is well worth the effort.
August 13, 2005