Tender on Private Cloud - part three
In the third and final part, I will describe a few points in the design of the architectural complex FlexPod, on which we have worked closely with Cisco and NetApp. There are things you should pay attention to.
Part one , second .
The Cisco UCS B200-M3 Blade server supports Intel Xeon E5-2600 processors and the obvious question is which processor you need to choose, combining performance and price. The customer did not specify what he means by 1 vcpu in physical equivalent - 1 core, 1 GHz, etc. Together with Cisco, we come to the next model. Based on the tests http://www.spec.org/cpu2006/ , we find a server with a ~ 1GHz processor and one heart. The result of his test is ~ 8-10. We take this result for 1 vcpu. Find the test result for the blade servers B200-M3 and compare. Example:
')
From the resulting, we find the ratio core / vcpu:
The difference in the ratio of vcpu / core is small, with a significant difference in price. The vcpu / core ratio of 4.5 is quite pessimistic today. You can raise it to 6 and get 96 vcpu for each blade.
The vcpu / core ratio is directly related to the cost of a virtual machine: the more machines spinning on one blade, the cheaper it is. Obviously, the cost of a virtual machine with a bronze level is cheaper than with a platinum one. In part, this is realized through the software and hardware complex. Thus, depending on the service (bronze, silver, gold or platinum), different vcpu / core ratios can be proposed. For example, for bronze to offer a high ratio, and vice versa, lower for platinum.
Currently, the FlexPod-certified model is the 6248UP 48-port series. I will cite the selected FlexPod architecture again.
Let 8 ports for each FI go uplink in the direction of Nexus and MDS. There are 40 ports for connecting the chassis.
2 Fabric Extender with 4 or 8 ports 10 GE can be installed in the chassis. In total complexity we get 80 Gb / s or 160 Gb / s on the chassis.
Depending on the required number of Fex ports per chassis, the maximum FlexPod configuration is determined:
Considering that there are 8 server blades on the chassis, it is easy to get the maximum number on the FlexPod and the total number of vcpu.
The new Fabric Interconnect 6296UP with 96 ports has already been released, which would expand the possibilities, but so far it has not received the FlexPod certification. True, we have excluded this model, for financial reasons:
If you immediately buy an expensive model Fabric Interconnect 6296UP, then this will affect the price of the virtual machine, which will be unreasonably high. Especially if more than 1000 cars will not work.
All the above elements were taken into account in order to propose not only a technically interesting solution, but also so that the total cost of a virtual machine would be reasonable.
Part one , second .
Blade server
The Cisco UCS B200-M3 Blade server supports Intel Xeon E5-2600 processors and the obvious question is which processor you need to choose, combining performance and price. The customer did not specify what he means by 1 vcpu in physical equivalent - 1 core, 1 GHz, etc. Together with Cisco, we come to the next model. Based on the tests http://www.spec.org/cpu2006/ , we find a server with a ~ 1GHz processor and one heart. The result of his test is ~ 8-10. We take this result for 1 vcpu. Find the test result for the blade servers B200-M3 and compare. Example:
- Cisco UCS B200 M3 (2 x Intel Xeon E5-2665, 2.40 GHz): 586 => 586/8 = 73.2 vcpu
- Cisco UCS B200 M3 (2 x Intel Xeon E5-2680, 2.70 GHz): 639 => 639/8 = 79.8 vcpu
')
From the resulting, we find the ratio core / vcpu:
- Intel Xeon E5-2665, 2.40 GHz: 73.2 / 16 = 4.5 vcpu for each core
- Intel Xeon E5-2680, 2.70 GHz: 79.8 / 16 = 4.9 vcpu for each core
The difference in the ratio of vcpu / core is small, with a significant difference in price. The vcpu / core ratio of 4.5 is quite pessimistic today. You can raise it to 6 and get 96 vcpu for each blade.
The vcpu / core ratio is directly related to the cost of a virtual machine: the more machines spinning on one blade, the cheaper it is. Obviously, the cost of a virtual machine with a bronze level is cheaper than with a platinum one. In part, this is realized through the software and hardware complex. Thus, depending on the service (bronze, silver, gold or platinum), different vcpu / core ratios can be proposed. For example, for bronze to offer a high ratio, and vice versa, lower for platinum.
Fabric interconnect
Currently, the FlexPod-certified model is the 6248UP 48-port series. I will cite the selected FlexPod architecture again.
Let 8 ports for each FI go uplink in the direction of Nexus and MDS. There are 40 ports for connecting the chassis.
2 Fabric Extender with 4 or 8 ports 10 GE can be installed in the chassis. In total complexity we get 80 Gb / s or 160 Gb / s on the chassis.
Depending on the required number of Fex ports per chassis, the maximum FlexPod configuration is determined:
- 1 x 10 Gb / s: 40 chassis
- 2 x 10 Gb / s: 20 chassis
- 4 x 10 Gb / s: 10 chassis
- 8 x 10 Gb / s: 5 chassis
Considering that there are 8 server blades on the chassis, it is easy to get the maximum number on the FlexPod and the total number of vcpu.
The new Fabric Interconnect 6296UP with 96 ports has already been released, which would expand the possibilities, but so far it has not received the FlexPod certification. True, we have excluded this model, for financial reasons:
- a solid order was received for only 1000 virtual machines
- there could potentially be several thousand virtual machines
If you immediately buy an expensive model Fabric Interconnect 6296UP, then this will affect the price of the virtual machine, which will be unreasonably high. Especially if more than 1000 cars will not work.
All the above elements were taken into account in order to propose not only a technically interesting solution, but also so that the total cost of a virtual machine would be reasonable.
Source: https://habr.com/ru/post/150480/
All Articles