Is it really cheap to integrate storage and calculation?

Recently, I visited an internet financial customer. The customer said that the current distributed architecture uses a database provided by an Internet, with 1 master and 2 slave. The hardware is a server and several local disks, with a scale of thousands of units, the annual capacity growth needs to be expanded. Due to the limited IO Capacity and Capacity of local disks, server resources need to be increased. However, the CPU utilization of database servers is very low, less than 10%.

The customer thinks that why the CPU resources must be expanded to expand the capacity? The CPU is very idle now, and the CPU is much more expensive than SSD when the two are bound!

For simple calculation, if you expand the capacity of 2TB, you need to add 3 servers with 1 master and 2 slave servers. This is 300000, and the cost is no less than 150000/TB! It is also necessary to increase the software license for the database, which is even worse. The capacity increases every year, but the business pressure does not increase significantly. How expensive is SSD disk pulling CPU into the water? Why can't it be separated?

Now we need to reduce costs and increase efficiency. You can create a separate SSD expansion plan, which can save a lot of costs. In addition, some servers are now equipped with 4 3.84TB disks with a bare capacity of 15TB. However, the actual capacity requirement for single data shard is only 1TB, and the capacity utilization rate is also very low. What percentage? Is there any solution?

It seems that Internet financial users also realize that the SSD disk of the database is tied to the CPU of the server. To expand the capacity, the CPU needs to be expanded synchronously. The ratio of the two must not be accurately calculated, and the resource utilization rate is low, resulting in serious waste, the cost is actually much higher.

In fact, in the past 30 years, the IOE technology architecture is that database servers are attached to external enterprise storage, storage and computing are separated, IO capabilities are uninstalled to storage, CPU resource consumption is reduced, capacity expansion is achieved, and disk expansion is achieved, if the CPU is insufficient, add the computing node. The two nodes are flexibly configured and are not bound. Moreover, computing nodes are stateless, so failover is fast. In terms of management, the management of resources must be gradually refined, not extensive. However, in recent years, some manufacturers have extensively introduced this multi-replica master-slave architecture, which has become the trend of the distributed architecture of the so-called & ldquo; Public praise & rdquo;. In fact, if you look at it carefully, is this really the case? Internet financial users who have not been in the early stage are also reflecting on how to achieve cost reduction and efficiency improvement through structural improvement.

Saving is a kind of virtue, so is architecture planning. It is very elementary to pile up multiple copies of resources. IOE architecture is a peering replacement, and it is already awesome to replace it. If the architecture is overturned, isn't it going back to the 1990 s?

The wisdom of Silicon Valley bosses over the past 30 years has proved to be successful. There are professional companies in every field. In common sense, professional companies or talents can do professional things.

Wenhang Source: Bi Xu said