Shulou(Shulou.com)06/03 Report--

The following is to analyze and describe the complete appearance of each kind of load balancing, hoping to bring some help to you in practical application. There are many things involved in load balancing, there are not many theories, and there are many books on the Internet. Today, we will use the accumulated experience in the industry to make an answer.

What is "load balancing"?

As shown in the figure, the process of converging traffic by an independent unified entrance and then doing secondary distribution is "load balancing", which, like the "distributed system", is "divide and conquer".

If you are used to using some navigation software when driving, we will find that there is an upper limit on the number of recommended routes in the navigation software, such as 3 or 5. Therefore, in fact, in essence, it also plays a similar role of "load balancing", because if you can only take the unobstructed route of Top3, the route with serious natural congestion cannot be recommended to you, so that the pressure of traffic flow is shared to the relatively idle route.

The same is true in software systems, in order to avoid uneven traffic sharing, resulting in excessive load on local nodes (such as tight CPU, etc.), so an independent unified entrance is introduced to do similar "navigation" work. However, the difference between "load balancing" in software system and navigation is that navigation is a flexible strategy, which ultimately requires users to make a choice, while the former is different.

Behind the balance is the strategy at work, and behind the strategy is made up of some algorithms or logic. For example, the algorithm in navigation belongs to the category of "path planning", which is subdivided into "static path planning" and "dynamic path planning". Moreover, there are a variety of specific algorithms under different branches, such as Dijikstra, A* and so on. Similarly, load balancing in software systems, there are many algorithms or logic to support these strategies, coincidentally, there are also static and dynamic.

Second, the diagram of commonly used "load balancing" strategy.

Here's a list of the five most common strategies in your daily work.

01 polling

This is the most commonly used and simplest strategy, evenly distributed by everyone, once per person. The approximate code is as follows.

Int globalIndex = 0; / / Note that it is a global variable, not a local variable. Try {return servers [globalIndex];} finally {globalIndex++; if (globalIndex = = 3) globalIndex = 0;} 02 weighted polling

On the basis of polling, a concept of weight is added. Weight is a generalized concept, which can be reflected in any way. In essence, it is a thought that those who can do more work. For example, you can configure different weights according to the performance differences of the host. The approximate code is as follows.

MatchedIndex = -; total = (I =; I

< servers.Length; i++){ servers[i].cur_weight += servers[i].weight; total += servers[i].weight; (matchedIndex == - || servers[matchedIndex].cur_weight < servers[i].cur_weight) { matchedIndex = i; }}servers[matchedIndex].cur_weight -= total; servers[matchedIndex]; 这段代码的过程如下图的表格。"()"中的数字就是自增数，代码中的cur_weight。 值得注意的是，加权轮询本身还有不同的实现方式，虽说最终的比例都是2：1：2。但是在请求送达的先后顺序上可以所有不同。比如「5-4，3，2-1」和上面的案例相比，最终比例是一样的，但是效果不同。「5-4，3，2-1」更容易产生并发问题，导致服务端拥塞，且这个问题随着权重数字越大越严重。例子：10：5：3的结果是「18-17-16-15-14-13-12-11-10-9，8-7-6-5-4，3-2-1」 03 最少连接数 这是一种根据实时的负载情况，进行动态负载均衡的方式。维护好活动中的连接数量，然后取最小的返回即可。大致的代码如下。 var matchedServer = servers.orderBy(e =>

E.active_conns) .first (); matchedServer.active_conns + = 1 th return matchedServer;// also needs to subtract 1 from active_conns when the connection is closed. 04 fastest response

This is also a dynamic load balancing strategy, its essence is to allocate according to the response of each node in the past period of time, the faster the response, the more allocation. There are also many ways to operate, and the above figure can be understood as recording the average time spent on requests in the most recent period, combined with the previous "weighted polling", so it is equivalent to 2:1:3 weighted polling.

Digression: generally speaking, the delay under the computer room is basically the same as that under the computer room, and the difference in response time is mainly in the processing capacity of the service. If it is used in some request processing across regions (for example, Zhejiang-> Shanghai or Zhejiang-> Beijing), in most cases, the timed "ping" method will be used to obtain the delay, because the L3 forwarding of OSI makes the data cleaner and more accurate.

05 Hash method

The load balancing of the hash method is different from the previous ones in that its result is determined by the client. A certain identity brought by the client is scattered and apportioned by a standardized hash function.

The hash function in the figure above uses the simplest and roughest "remainder method".

Aside from the topic: in addition to the remainder of the hash function, there are other hash functions, such as "changing base", "folding", "square middle method" and so on.

In addition, the parameter to be solved can be arbitrary, as long as it is finally converted into an integer to participate in the operation. The most commonly used method is to use the source ip address as a parameter to ensure that the same client requests fall on the same CVM as much as possible.

Third, the advantages, disadvantages and applicable scenarios of common "load balancing" strategies.

We know that there is no perfect thing, and so is the load balancing strategy. The most commonly used strategies listed above also have their own advantages and disadvantages and applicable scenarios, which I have sorted out a bit, as follows.

The reason why these load balancing algorithms are commonly used is also because it is simple. If you want a better effect, you must need higher complexity. For example, you can use a combination of simple strategies, or comprehensive evaluation through more dimensional data sampling, or even based on prediction algorithms after data mining.

Fourth, use "health detection" to ensure high availability

No matter what kind of strategy it is, it is inevitable to encounter machine failure or program failure. Therefore, in order to ensure that load balancing can work better, we also need to combine some "health detection" mechanisms. Regularly detect whether the server can still be connected and whether the response is slower than expected. If the node is in an "unavailable" state, it needs to be temporarily removed from the waiting list to improve availability. There are three commonly used ways of "health detection".

01 HTTP probe

Use Get/Post to request a fixed URL on the server to determine whether the returned content is in line with expectations. Generally use the Http status code, the content in response to judge.

02 TCP probe

Based on Tcp's three-way handshake mechanism to detect the specified IP + port. Best practices can learn from Aliyun's SLB mechanism, as shown in the following figure.

▲

It is worth noting that in order to release the connection as soon as possible, follow the RST immediately after the end of the three-way handshake to break the TCP connection.

03 UDP probe

There may be UDP protocols used by some applications. Under this protocol, the specified IP + port can be detected by message. Best practices can also learn from Aliyun's SLB mechanism, as shown in the following figure.

The way to determine the result is: if the server does not return any information, the default normal state. Otherwise, an error message of ICMP will be returned.

After reading the above analysis and describing the complete appearance of each kind of load balancing, if there is anything else you need to know, you can find out what you are interested in in the industry information or find our professional and technical engineers for answers. Technical engineers have more than ten years of experience in the industry.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.