25G Ethernet VS 40G Ethernet, which is Right for You?

25G and 40G Ethernet are the Next Generation data center rate options, how would you choose?  Our answer is 25G Ethernet. Why 25G Ethernet is chosen as the next generation of the mainstream rate, in short: more efficient, more scalable, more cost-effective.

Why 25G Ethernet?

1. Although the rate of 40G is faster than 25G, we know that the current 40G is achieved through 4x10G, a generation behind the 25G. Assuming that a swap chip has 128 devices that are LANs, the entire bandwidth and switching capacity at 25G is the largest, while the number of servers connected to the 25G rate and the exchange capacity is also optimal.
2. 25G can be more easily upgraded to 50G and 100G, 25G technology before the emergence of 100G is achieved through 10 × 10G and application, now for 25G technology, 4 × 25G can achieve 100G. If you use PAM4 level, a single 25G LANE can achieve 50Gb/s data transmission.
3. The consumption and cost of individual Gbit have decreased relative to 10G. First, power consumption, we have done a comparison, the switch is filled with AOC and optical modules, serpentine full load case, 25G switch power consumption is about 1.5 times of the 10G switch. The 25G consumption/Gbit is 0.6 times of the 10G. Regarding for the cost, we expect that the cost of the 25G system will be about twice times the cost of the 10G system, compared with the costs of the AOC and the optical module at the end of the year. So the 25G cost/Gbit is 0.8 times that of 10G, and we believe that this number will decrease with the application of 25G next year. Power consumption, the 25G switch power consumption is approximately 1.5 times the power consumption of 10G switch under full load condition.
4. 25G Ecological Maturity. The 25G switch in the application of data center is a real application of engineering practice, so we need 25G ecological upstream and downstream of all aspects of maturity, products can be high-volume production, 25G switch and 25G-related engineering technology has matured applications or will be a mature application. The above is the choice of 25G external conditions, the other is to meet user needs. The data traffic for some Hadoop clusters started last year, with the rapid development of cloud computing and AI, increasing demand for network latency and bandwidth, we need upgrading from 10G networks to 25G networks to provide higher bandwidth.

Based on the above reasons, we choose 25G as the mainstream rate of the data center. ODCC’s research on 25G began in 2015 and October 2015, when a network group was set up, and a number of projects were presented, one of which was the 25G switch project. September 2016 ODCC Summit released 25G ToR Switch specification book V1.0 version. This specification details the hardware architecture and core design of the 25G Tor switch recommended by ODCC. It is worth mentioning that this specification book gives two different specifications for developers to refer to. 2017 is 25G in the year of the ODCC outbreak, April 2017 we released the 25G AOC and 100G Open optical transceiver technical specification.

25G TOR design principles

The first is to solve the 10G switch pain points, the first pain point is unified software hardware processing. Now there are a lot of 10 Gigabit switch manufacturers on the market, the use of the CPU, and some manufacturers use MIPS architecture CPU, and some use PowerPC architecture CPU, while the hardware system driver does not control. CPU architecture and hardware systems are not unified cause R & D OS time to adapt, the network management caused some trouble. The second is the I2C problem. I2C bus in the switch system has a wide range of applications, SFP28 AOC, optical modules, PSU, FAN management, mostly using the I2C bus. For the pluggable optical transceiver module, the plug-in optical transceiver module, a certain probability of I2C bus lock will appear, if not in this way the I2C bus to do isolation, which will affect the entire I2C bus. Resulting in the I2C bus is not available, to bring some inconvenience to the maintenance. The third pain point is not enough means of diagnosis is not rich, the less problem reported, if online switches down, we can not remote diagnosis, cannot be remote recovery, we must notify the scene of the manual recovery of manual, Resources are also a waste. If the switch fails to restart, we restart the type of restart and the reasons for the record is relatively small, the problem of failure warning has not been achieved.