You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@dolphinscheduler.apache.org by Yichao Yang <10...@qq.com> on 2020/07/19 03:13:10 UTC
Re： [DISCUSS]Load balancing weight
Hi,
I got your point.
I agree with your about the all algorithms you mentioned should be measured by weight. But I think we should keep the original algorithm which is not based on weight selection. and the algorithm based on weight should appear in a new algorithm.
In my opinion. Actually in our production environment, there are few differences in the performance of the machines as you mentioned. In most cases, the machine performance and configuration are consistent in our production environment.
So there are two scenarios for this weight based algorithm. The first one is the one you mentioned. The second may occur when a user wants to use half of the performance of machine A, rather than full use of machine A.
So is it better to keep the original algorithm and also provide those new selection algorithms based on weight?
Best,
Yichao Yang
------------------ 原始邮件 ------------------
发件人: "dev" <acmrabbit@gmail.com>;
发送时间: 2020年7月18日(星期六) 晚上10:12
收件人: "dev"<dev@dolphinscheduler.apache.org>;
主题: Re: [DISCUSS]Load balancing weight
Let me give an example. It might be better. Suppose my computer A has twice
the processing power of computer B. Then I have three tasks. According to
the polling algorithm, A will handle two and B will handle one. Isn't our
purpose of load balancing to make each machine work evenly?
Best wishes！
CalvinKirs>
On 2020/07/18 13:56:38, lidong dai <d...@gmail.com> wrote:
> I have a little confused about adding a weight to what you said, like>
> assign tasks to machines in turn, there is normally no weights>
>
>
>
> Best Regards>
> --------------->
> DolphinScheduler(Incubator) PPMC>
> Lidong Dai 代立冬>
> dailidong66@gmail.com>
> --------------->
>
>
> CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
>
> > Hi, everyone. I’m currently designing weight-based load balancing. My>
> > point of view is that all load balancing needs to be done based on
weights,>
> > because load actually means that machines with good performance are>
> > expected to handle more tasks. Therefore, weights should be Reflected
in>
> > each load balancing algorithm, it should be a default attribute in
load>
> > balancing.>
> >>
> > Hope to get your suggestions.>
> >>
> >>
> > Best wishes！>
> > CalvinKirs>
> >>
> >>
>
Re： Re： [DISCUSS]Load balancing weight
Posted by Yichao Yang <10...@qq.com>.
Hi all,
Is there anyone have some suggestions about this discuss?
Best,
Yichao Yang
------------------ 原始邮件 ------------------
发件人: "dev" <acmrabbit@gmail.com>;
发送时间: 2020年7月19日(星期天) 晚上8:29
收件人: "dev"<dev@dolphinscheduler.apache.org>;
主题: Re: Re： [DISCUSS]Load balancing weight
I think the smooth weighted polling algorithm is an enhancement to the
polling algorithm, and they belong to the containment relationship. Perhaps
we should listen to the opinions of other partners in the community?
我认为平滑加权轮询算法是对轮询算法的增强，它们属于包含关系。一开始最早的版本我写的是加权轮询，后面又改成了平滑加权轮询。他们意义上的确有差异，但同样属于包含关系，类似的功能增强都如此，我们不应该在功能增强的时候还应该保留老的文件。
也许我们应该听取社区中其他伙伴的意见。
Yichao Yang <1048262223@qq.com>于2020年7月19日 周日下午6:24写道：
> Hi,
>
>
> I got your point. And from the point of the code logic and the functions
> that can be used after code implementation, I agree that the weighted
> polling algorithm is indeed applicable to scenarios without setting weights.
>
>
> But in terms of module division and architecture design, the weight will
> be coupled into the existing simple round robin algorithm. In fact, the
> weighted round robin&nbsp;algorithm is also a separate algorithm, which is
> different from the simple&nbsp;round robin&nbsp;algorithm in terms of
> design[1].
>
>
> What I want to express in my last email[2] is that from the perspective of
> module division and enhancement of module clarity, Is it better
> to&nbsp;introduce weighted rotation training and other algorithms into new
> algorithms instead of adding weight logic to the previous algorithms?
>
>
>
> --------------
>
>
> 我明白你的观点，从代码逻辑以及实现后能展现出来的功能角度上，我是同意加权轮询算法确实是适用于没有设置权重的场景的。
>
>
>
> 但是，在模块划分上和架构设计上，其实是把权重这部分给耦合进了现有的简单轮训算法中了。加权轮训算法其实也是一种单独的算法，和简单轮训算法在设计角度和功能上是不一样的[1]。
>
>
>
> 我在上一份邮件中表达的意思是说，从模块划分，增强模块清晰度的角度，我们把加权轮训等算法以新的算法引入，而不去在以前的算法上加入权重的逻辑是不是会更好点呢？
>
>
> [1]&nbsp;https://blog.csdn.net/bohu83/article/details/79669051
> [2]&nbsp;
> https://lists.apache.org/thread.html/r05005385cd0a52be8e601fd1e76a8dacbf3a24c679a3a435eb1d9c8a%40%3Cdev.dolphinscheduler.apache.org%3E
>
>
> If you have any question or suggetion, welcome to put forward~
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------&nbsp;原始邮件&nbsp;------------------
> 发件人:
> "dev"
> <
> acmrabbit@gmail.com&gt;;
> 发送时间:&nbsp;2020年7月19日(星期天) 下午5:31
> 收件人:&nbsp;"dev"<dev@dolphinscheduler.apache.org&gt;;
>
> 主题:&nbsp;Re: Re： [DISCUSS]Load balancing weight
>
>
>
> Hi, I think you have misunderstood. In fact, the weighted smooth polling
> algorithm is also suitable for scenarios where no weight is set.
>
> Best&nbsp; wishes！
> CalvinKirs
>
> On 2020/07/19 03:13:10, "Yichao Yang" <1048262223@qq.com&gt; wrote:
> &gt; Hi,
> &gt;
> &gt;
> &gt; I got your point.
> &gt;
> &gt;
> &gt; I agree with your about the all algorithms you mentioned should be
> measured by weight. But I think we should keep the original algorithm which
> is not based on weight selection. and the algorithm based on weight should
> appear in a new algorithm.
> &gt;
> &gt;
> &gt; In my opinion. Actually in our production environment, there are few
> differences in the performance of the machines as you mentioned. In most
> cases, the machine performance and configuration are consistent in our
> production environment.
> &gt;
> &gt;
> &gt; So there are two scenarios for this weight based algorithm. The first
> one is the one you mentioned. The second may occur when a user wants to use
> half of the performance of machine A, rather than full use of machine A.
> &gt;
> &gt;
> &gt; So is it better to keep the original algorithm and also provide those
> new selection algorithms based on weight?
> &gt;
> &gt;
> &gt;
> &gt;
> &gt;
> &gt; Best,
> &gt; Yichao Yang
> &gt;
> &gt;
> &gt;
> &gt;
> &gt; ------------------&nbsp;原始邮件&nbsp;------------------
> &gt;
> 发件人:&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
> "dev"&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
> <acmrabbit@gmail.com&gt;;
> &gt; 发送时间:&nbsp;2020年7月18日(星期六) 晚上10:12
> &gt; 收件人:&nbsp;"dev"<dev@dolphinscheduler.apache.org&gt;;
> &gt;
> &gt; 主题:&nbsp;Re: [DISCUSS]Load balancing weight
> &gt;
> &gt;
> &gt;
> &gt; Let me give an example. It might be better. Suppose my computer A has
> twice
> &gt; the processing power of computer B. Then I have three tasks.
> According to
> &gt; the polling algorithm, A will handle two and B will handle one. Isn't
> our
> &gt; purpose of load balancing to make each machine work evenly?
> &gt;
> &gt; Best&nbsp; wishes！
> &gt; &nbsp;CalvinKirs&gt;
> &gt;
> &gt; On 2020/07/18 13:56:38, lidong dai <d...@gmail.com&gt; wrote:
> &gt; &gt; I have a little confused about adding a weight to what you said,
> like&gt;
> &gt; &gt; assign tasks to machines in turn, there is normally no
> weights&gt;
> &gt; &gt;
> &gt; &gt;
> &gt; &gt;
> &gt; &gt; Best Regards&gt;
> &gt; &gt; ---------------&gt;
> &gt; &gt; DolphinScheduler(Incubator) PPMC&gt;
> &gt; &gt; Lidong Dai 代立冬&gt;
> &gt; &gt; dailidong66@gmail.com&gt;
> &gt; &gt; ---------------&gt;
> &gt; &gt;
> &gt; &gt;
> &gt; &gt; CalvinKirs <ac...@163.com&gt; 于2020年7月18日周六 下午5:55写道：&gt;
> &gt; &gt;
> &gt; &gt; &gt; Hi, everyone. I’m currently designing weight-based load
> balancing. My&gt;
> &gt; &gt; &gt; point of view is that all load balancing needs to be done
> based on
> &gt; weights,&gt;
> &gt; &gt; &gt; because load actually means that machines with good
> performance are&gt;
> &gt; &gt; &gt; expected to handle more tasks. Therefore, weights should be
> Reflected
> &gt; in&gt;
> &gt; &gt; &gt; each load balancing algorithm, it should be a default
> attribute in
> &gt; load&gt;
> &gt; &gt; &gt; balancing.&gt;
> &gt; &gt; &gt;&gt;
> &gt; &gt; &gt; Hope to get your suggestions.&gt;
> &gt; &gt; &gt;&gt;
> &gt; &gt; &gt;&gt;
> &gt; &gt; &gt; Best&nbsp; wishes！&gt;
> &gt; &gt; &gt; CalvinKirs&gt;
> &gt; &gt; &gt;&gt;
> &gt; &gt; &gt;&gt;
> &gt; &gt;
Re: Re： [DISCUSS]Load balancing weight
Posted by Kris Calvin <ac...@gmail.com>.
I think the smooth weighted polling algorithm is an enhancement to the
polling algorithm, and they belong to the containment relationship. Perhaps
we should listen to the opinions of other partners in the community?
我认为平滑加权轮询算法是对轮询算法的增强，它们属于包含关系。一开始最早的版本我写的是加权轮询，后面又改成了平滑加权轮询。他们意义上的确有差异，但同样属于包含关系，类似的功能增强都如此，我们不应该在功能增强的时候还应该保留老的文件。
也许我们应该听取社区中其他伙伴的意见。
Yichao Yang <10...@qq.com>于2020年7月19日 周日下午6:24写道：
> Hi,
>
>
> I got your point. And from the point of the code logic and the functions
> that can be used after code implementation, I agree that the weighted
> polling algorithm is indeed applicable to scenarios without setting weights.
>
>
> But in terms of module division and architecture design, the weight will
> be coupled into the existing simple round robin algorithm. In fact, the
> weighted round robin algorithm is also a separate algorithm, which is
> different from the simple round robin algorithm in terms of
> design[1].
>
>
> What I want to express in my last email[2] is that from the perspective of
> module division and enhancement of module clarity, Is it better
> to introduce weighted rotation training and other algorithms into new
> algorithms instead of adding weight logic to the previous algorithms?
>
>
>
> --------------
>
>
> 我明白你的观点，从代码逻辑以及实现后能展现出来的功能角度上，我是同意加权轮询算法确实是适用于没有设置权重的场景的。
>
>
>
> 但是，在模块划分上和架构设计上，其实是把权重这部分给耦合进了现有的简单轮训算法中了。加权轮训算法其实也是一种单独的算法，和简单轮训算法在设计角度和功能上是不一样的[1]。
>
>
>
> 我在上一份邮件中表达的意思是说，从模块划分，增强模块清晰度的角度，我们把加权轮训等算法以新的算法引入，而不去在以前的算法上加入权重的逻辑是不是会更好点呢？
>
>
> [1] https://blog.csdn.net/bohu83/article/details/79669051
> [2]
> https://lists.apache.org/thread.html/r05005385cd0a52be8e601fd1e76a8dacbf3a24c679a3a435eb1d9c8a%40%3Cdev.dolphinscheduler.apache.org%3E
>
>
> If you have any question or suggetion, welcome to put forward~
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------ 原始邮件 ------------------
> 发件人:
> "dev"
> <
> acmrabbit@gmail.com>;
> 发送时间: 2020年7月19日(星期天) 下午5:31
> 收件人: "dev"<dev@dolphinscheduler.apache.org>;
>
> 主题: Re: Re： [DISCUSS]Load balancing weight
>
>
>
> Hi, I think you have misunderstood. In fact, the weighted smooth polling
> algorithm is also suitable for scenarios where no weight is set.
>
> Best wishes！
> CalvinKirs
>
> On 2020/07/19 03:13:10, "Yichao Yang" <1048262223@qq.com> wrote:
> > Hi,
> >
> >
> > I got your point.
> >
> >
> > I agree with your about the all algorithms you mentioned should be
> measured by weight. But I think we should keep the original algorithm which
> is not based on weight selection. and the algorithm based on weight should
> appear in a new algorithm.
> >
> >
> > In my opinion. Actually in our production environment, there are few
> differences in the performance of the machines as you mentioned. In most
> cases, the machine performance and configuration are consistent in our
> production environment.
> >
> >
> > So there are two scenarios for this weight based algorithm. The first
> one is the one you mentioned. The second may occur when a user wants to use
> half of the performance of machine A, rather than full use of machine A.
> >
> >
> > So is it better to keep the original algorithm and also provide those
> new selection algorithms based on weight?
> >
> >
> >
> >
> >
> > Best,
> > Yichao Yang
> >
> >
> >
> >
> > ------------------ 原始邮件 ------------------
> >
> 发件人:
> "dev"
> <acmrabbit@gmail.com>;
> > 发送时间: 2020年7月18日(星期六) 晚上10:12
> > 收件人: "dev"<dev@dolphinscheduler.apache.org>;
> >
> > 主题: Re: [DISCUSS]Load balancing weight
> >
> >
> >
> > Let me give an example. It might be better. Suppose my computer A has
> twice
> > the processing power of computer B. Then I have three tasks.
> According to
> > the polling algorithm, A will handle two and B will handle one. Isn't
> our
> > purpose of load balancing to make each machine work evenly?
> >
> > Best wishes！
> > CalvinKirs>
> >
> > On 2020/07/18 13:56:38, lidong dai <d...@gmail.com> wrote:
> > > I have a little confused about adding a weight to what you said,
> like>
> > > assign tasks to machines in turn, there is normally no
> weights>
> > >
> > >
> > >
> > > Best Regards>
> > > --------------->
> > > DolphinScheduler(Incubator) PPMC>
> > > Lidong Dai 代立冬>
> > > dailidong66@gmail.com>
> > > --------------->
> > >
> > >
> > > CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
> > >
> > > > Hi, everyone. I’m currently designing weight-based load
> balancing. My>
> > > > point of view is that all load balancing needs to be done
> based on
> > weights,>
> > > > because load actually means that machines with good
> performance are>
> > > > expected to handle more tasks. Therefore, weights should be
> Reflected
> > in>
> > > > each load balancing algorithm, it should be a default
> attribute in
> > load>
> > > > balancing.>
> > > >>
> > > > Hope to get your suggestions.>
> > > >>
> > > >>
> > > > Best wishes！>
> > > > CalvinKirs>
> > > >>
> > > >>
> > >
Re: Re: [DISCUSS]Load balancing weight
Posted by "qiaozhanwei@outlook.com" <qi...@outlook.com>.
agree.
In my opinion . there are two control strategies for load balancing
1. assign weight .
2. calculate the weight of various indicators，for retry、response time and so on
weights can be embedded in polling and random algorithms . otherwise, there will be many load balancing algorithms, and the weight-based algorithm itself relies on polling and random algorithms, it does not exist independently
同意 CalvinKirs 观点
我个人认为对于负载均衡，有两种控制策略
1，手动的设置一些权重，根据一些情况，比如说机器情况、业务情况等等
2，根据很多指标来计算一些权重，比如说根据cpu、内存、和负载线性加权，已经实现了。比如说失败重试、响应时间策略进行负载均衡
权重的设置可以嵌入到轮询和随机算法中，否则就会有权重的单独的负载均衡算法，我个人认为权重是依赖于轮询和随机算法的，不是可以独立存在的
Thx
―――――――――――――
DolphinScheduler(Incubator) PPMC
Zhanwei Qiao 乔占卫
qiaozhanwei@outlook.com
From: CalvinKirs
Date: 2020-07-29 21:12
To: dev@dolphinscheduler.apache.org
Subject: Re: [DISCUSS]Load balancing weight
In addition, in my opinion, adding weights is only to improve load balancing, and simple polling and weighted polling are actually polling. Just imagine, when JIT needs to consider, when the computer hardware configuration is different (or temporarily add a machine), do we still need users to switch from simple polling to weighted polling?
另外，在我看来，增加权重也只是完善负载均衡，简单轮询和加权轮询其实本质依然是轮询。试想一下，当jit需要考虑的时候，当计算机硬件配置不一样的时候（或者临时加机器），难道我们还需要用户从简单轮询切换到加权轮询吗？
Best wishes！
CalvinKirs
On 07/29/2020 20:45，CalvinKirs<ac...@163.com> wrote：
Hello, thank you very much for your reply, but I still think it is unnecessary to do so. Weight belongs to the basis of load balancing. It is an indispensable part of load balancing. Real load balancing cannot be achieved by ignoring weight. Another point is that weight is composed of multiple attributes, such as a performance of the machine. In addition, we may also need Consider the weight impact brought by the optimization of JIT. I think this is an enhancement, not a new algorithm. Although it is very easy for me to add two new files, it is completely unnecessary. For users, there are a few more choices, especially when A enhances B but can do the same work as B ( That is to ignore the weight), this choice is very tasteless. Addition is always easier than subtraction, and when we want to do subtraction, it may be much more difficult.
hello,非常感谢您的回复，但是我依然认为没有必要这么做。权重属于负载均衡的基础，它是负载均衡不可缺少的一部分，忽略权重并不能够达到真正的负载均衡，另外一点，权重是由多个属性组成的，比如机器的一个性能，另外我们可能还需要考虑JIT的优化工作所带来的权重影响。我认为这属于增强，而不是新的算法。虽然对我来讲，新加两份文件会非常容易，但那完全没必要，对于用户来讲，多几个选择，尤其是A增强了B但是完全可以做到和B一样的工作的时候（即忽略权重），这种选择很鸡肋。做加法永远比做减法容易，而当我们想要做减法的时候，难度可能会大很多。
Best wishes！
CalvinKirs
On 07/28/2020 19:18，lidong dai<da...@gmail.com> wrote：
I also think as a new algorithm will be better
Best Regards
---------------
DolphinScheduler(Incubator) PPMC
Lidong Dai 代立冬
dailidong66@gmail.com
---------------
Yichao Yang <10...@qq.com> 于2020年7月19日周日 下午6:24写道：
Hi,
I got your point. And from the point of the code logic and the functions
that can be used after code implementation, I agree that the weighted
polling algorithm is indeed applicable to scenarios without setting weights.
But in terms of module division and architecture design, the weight will
be coupled into the existing simple round robin algorithm. In fact, the
weighted round robin algorithm is also a separate algorithm, which is
different from the simple round robin algorithm in terms of
design[1].
What I want to express in my last email[2] is that from the perspective of
module division and enhancement of module clarity, Is it better
to introduce weighted rotation training and other algorithms into new
algorithms instead of adding weight logic to the previous algorithms?
--------------
我明白你的观点，从代码逻辑以及实现后能展现出来的功能角度上，我是同意加权轮询算法确实是适用于没有设置权重的场景的。
但是，在模块划分上和架构设计上，其实是把权重这部分给耦合进了现有的简单轮训算法中了。加权轮训算法其实也是一种单独的算法，和简单轮训算法在设计角度和功能上是不一样的[1]。
我在上一份邮件中表达的意思是说，从模块划分，增强模块清晰度的角度，我们把加权轮训等算法以新的算法引入，而不去在以前的算法上加入权重的逻辑是不是会更好点呢？
[1] https://blog.csdn.net/bohu83/article/details/79669051
[2]
https://lists.apache.org/thread.html/r05005385cd0a52be8e601fd1e76a8dacbf3a24c679a3a435eb1d9c8a%40%3Cdev.dolphinscheduler.apache.org%3E
If you have any question or suggetion, welcome to put forward~
Best,
Yichao Yang
------------------ 原始邮件 ------------------
发件人:
"dev"
<
acmrabbit@gmail.com>;
发送时间: 2020年7月19日(星期天) 下午5:31
收件人: "dev"<dev@dolphinscheduler.apache.org>;
主题: Re: Re： [DISCUSS]Load balancing weight
Hi, I think you have misunderstood. In fact, the weighted smooth polling
algorithm is also suitable for scenarios where no weight is set.
Best wishes！
CalvinKirs
On 2020/07/19 03:13:10, "Yichao Yang" <1048262223@qq.com> wrote:
> Hi,
>
>
> I got your point.
>
>
> I agree with your about the all algorithms you mentioned should be
measured by weight. But I think we should keep the original algorithm which
is not based on weight selection. and the algorithm based on weight should
appear in a new algorithm.
>
>
> In my opinion. Actually in our production environment, there are few
differences in the performance of the machines as you mentioned. In most
cases, the machine performance and configuration are consistent in our
production environment.
>
>
> So there are two scenarios for this weight based algorithm. The first
one is the one you mentioned. The second may occur when a user wants to use
half of the performance of machine A, rather than full use of machine A.
>
>
> So is it better to keep the original algorithm and also provide those
new selection algorithms based on weight?
>
>
>
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------ 原始邮件 ------------------
>
发件人:
"dev"
<acmrabbit@gmail.com>;
> 发送时间: 2020年7月18日(星期六) 晚上10:12
> 收件人: "dev"<dev@dolphinscheduler.apache.org>;
>
> 主题: Re: [DISCUSS]Load balancing weight
>
>
>
> Let me give an example. It might be better. Suppose my computer A has
twice
> the processing power of computer B. Then I have three tasks.
According to
> the polling algorithm, A will handle two and B will handle one. Isn't
our
> purpose of load balancing to make each machine work evenly?
>
> Best wishes！
> CalvinKirs>
>
> On 2020/07/18 13:56:38, lidong dai <d...@gmail.com> wrote:
> > I have a little confused about adding a weight to what you said,
like>
> > assign tasks to machines in turn, there is normally no
weights>
> >
> >
> >
> > Best Regards>
> > --------------->
> > DolphinScheduler(Incubator) PPMC>
> > Lidong Dai 代立冬>
> > dailidong66@gmail.com>
> > --------------->
> >
> >
> > CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
> >
> > > Hi, everyone. I’m currently designing weight-based load
balancing. My>
> > > point of view is that all load balancing needs to be done
based on
> weights,>
> > > because load actually means that machines with good
performance are>
> > > expected to handle more tasks. Therefore, weights should be
Reflected
> in>
> > > each load balancing algorithm, it should be a default
attribute in
> load>
> > > balancing.>
> > >>
> > > Hope to get your suggestions.>
> > >>
> > >>
> > > Best wishes！>
> > > CalvinKirs>
> > >>
> > >>
> >
Re: [DISCUSS]Load balancing weight
Posted by CalvinKirs <ac...@163.com>.
In addition, in my opinion, adding weights is only to improve load balancing, and simple polling and weighted polling are actually polling. Just imagine, when JIT needs to consider, when the computer hardware configuration is different (or temporarily add a machine), do we still need users to switch from simple polling to weighted polling?
另外，在我看来，增加权重也只是完善负载均衡，简单轮询和加权轮询其实本质依然是轮询。试想一下，当jit需要考虑的时候，当计算机硬件配置不一样的时候（或者临时加机器），难道我们还需要用户从简单轮询切换到加权轮询吗？
Best wishes！
CalvinKirs
On 07/29/2020 20:45，CalvinKirs<ac...@163.com> wrote：
Hello, thank you very much for your reply, but I still think it is unnecessary to do so. Weight belongs to the basis of load balancing. It is an indispensable part of load balancing. Real load balancing cannot be achieved by ignoring weight. Another point is that weight is composed of multiple attributes, such as a performance of the machine. In addition, we may also need Consider the weight impact brought by the optimization of JIT. I think this is an enhancement, not a new algorithm. Although it is very easy for me to add two new files, it is completely unnecessary. For users, there are a few more choices, especially when A enhances B but can do the same work as B ( That is to ignore the weight), this choice is very tasteless. Addition is always easier than subtraction, and when we want to do subtraction, it may be much more difficult.
hello,非常感谢您的回复，但是我依然认为没有必要这么做。权重属于负载均衡的基础，它是负载均衡不可缺少的一部分，忽略权重并不能够达到真正的负载均衡，另外一点，权重是由多个属性组成的，比如机器的一个性能，另外我们可能还需要考虑JIT的优化工作所带来的权重影响。我认为这属于增强，而不是新的算法。虽然对我来讲，新加两份文件会非常容易，但那完全没必要，对于用户来讲，多几个选择，尤其是A增强了B但是完全可以做到和B一样的工作的时候（即忽略权重），这种选择很鸡肋。做加法永远比做减法容易，而当我们想要做减法的时候，难度可能会大很多。
Best wishes！
CalvinKirs
On 07/28/2020 19:18，lidong dai<da...@gmail.com> wrote：
I also think as a new algorithm will be better
Best Regards
---------------
DolphinScheduler(Incubator) PPMC
Lidong Dai 代立冬
dailidong66@gmail.com
---------------
Yichao Yang <10...@qq.com> 于2020年7月19日周日 下午6:24写道：
Hi,
I got your point. And from the point of the code logic and the functions
that can be used after code implementation, I agree that the weighted
polling algorithm is indeed applicable to scenarios without setting weights.
But in terms of module division and architecture design, the weight will
be coupled into the existing simple round robin algorithm. In fact, the
weighted round robin algorithm is also a separate algorithm, which is
different from the simple round robin algorithm in terms of
design[1].
What I want to express in my last email[2] is that from the perspective of
module division and enhancement of module clarity, Is it better
to introduce weighted rotation training and other algorithms into new
algorithms instead of adding weight logic to the previous algorithms?
--------------
我明白你的观点，从代码逻辑以及实现后能展现出来的功能角度上，我是同意加权轮询算法确实是适用于没有设置权重的场景的。
但是，在模块划分上和架构设计上，其实是把权重这部分给耦合进了现有的简单轮训算法中了。加权轮训算法其实也是一种单独的算法，和简单轮训算法在设计角度和功能上是不一样的[1]。
我在上一份邮件中表达的意思是说，从模块划分，增强模块清晰度的角度，我们把加权轮训等算法以新的算法引入，而不去在以前的算法上加入权重的逻辑是不是会更好点呢？
[1] https://blog.csdn.net/bohu83/article/details/79669051
[2]
https://lists.apache.org/thread.html/r05005385cd0a52be8e601fd1e76a8dacbf3a24c679a3a435eb1d9c8a%40%3Cdev.dolphinscheduler.apache.org%3E
If you have any question or suggetion, welcome to put forward~
Best,
Yichao Yang
------------------ 原始邮件 ------------------
发件人:
"dev"
<
acmrabbit@gmail.com>;
发送时间: 2020年7月19日(星期天) 下午5:31
收件人: "dev"<dev@dolphinscheduler.apache.org>;
主题: Re: Re： [DISCUSS]Load balancing weight
Hi, I think you have misunderstood. In fact, the weighted smooth polling
algorithm is also suitable for scenarios where no weight is set.
Best wishes！
CalvinKirs
On 2020/07/19 03:13:10, "Yichao Yang" <1048262223@qq.com> wrote:
> Hi,
>
>
> I got your point.
>
>
> I agree with your about the all algorithms you mentioned should be
measured by weight. But I think we should keep the original algorithm which
is not based on weight selection. and the algorithm based on weight should
appear in a new algorithm.
>
>
> In my opinion. Actually in our production environment, there are few
differences in the performance of the machines as you mentioned. In most
cases, the machine performance and configuration are consistent in our
production environment.
>
>
> So there are two scenarios for this weight based algorithm. The first
one is the one you mentioned. The second may occur when a user wants to use
half of the performance of machine A, rather than full use of machine A.
>
>
> So is it better to keep the original algorithm and also provide those
new selection algorithms based on weight?
>
>
>
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------ 原始邮件 ------------------
>
发件人:
"dev"
<acmrabbit@gmail.com>;
> 发送时间: 2020年7月18日(星期六) 晚上10:12
> 收件人: "dev"<dev@dolphinscheduler.apache.org>;
>
> 主题: Re: [DISCUSS]Load balancing weight
>
>
>
> Let me give an example. It might be better. Suppose my computer A has
twice
> the processing power of computer B. Then I have three tasks.
According to
> the polling algorithm, A will handle two and B will handle one. Isn't
our
> purpose of load balancing to make each machine work evenly?
>
> Best wishes！
> CalvinKirs>
>
> On 2020/07/18 13:56:38, lidong dai <d...@gmail.com> wrote:
> > I have a little confused about adding a weight to what you said,
like>
> > assign tasks to machines in turn, there is normally no
weights>
> >
> >
> >
> > Best Regards>
> > --------------->
> > DolphinScheduler(Incubator) PPMC>
> > Lidong Dai 代立冬>
> > dailidong66@gmail.com>
> > --------------->
> >
> >
> > CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
> >
> > > Hi, everyone. I’m currently designing weight-based load
balancing. My>
> > > point of view is that all load balancing needs to be done
based on
> weights,>
> > > because load actually means that machines with good
performance are>
> > > expected to handle more tasks. Therefore, weights should be
Reflected
> in>
> > > each load balancing algorithm, it should be a default
attribute in
> load>
> > > balancing.>
> > >>
> > > Hope to get your suggestions.>
> > >>
> > >>
> > > Best wishes！>
> > > CalvinKirs>
> > >>
> > >>
> >
Re: [DISCUSS]Load balancing weight
Posted by CalvinKirs <ac...@163.com>.
Hello, thank you very much for your reply, but I still think it is unnecessary to do so. Weight belongs to the basis of load balancing. It is an indispensable part of load balancing. Real load balancing cannot be achieved by ignoring weight. Another point is that weight is composed of multiple attributes, such as a performance of the machine. In addition, we may also need Consider the weight impact brought by the optimization of JIT. I think this is an enhancement, not a new algorithm. Although it is very easy for me to add two new files, it is completely unnecessary. For users, there are a few more choices, especially when A enhances B but can do the same work as B ( That is to ignore the weight), this choice is very tasteless. Addition is always easier than subtraction, and when we want to do subtraction, it may be much more difficult.
hello,非常感谢您的回复，但是我依然认为没有必要这么做。权重属于负载均衡的基础，它是负载均衡不可缺少的一部分，忽略权重并不能够达到真正的负载均衡，另外一点，权重是由多个属性组成的，比如机器的一个性能，另外我们可能还需要考虑JIT的优化工作所带来的权重影响。我认为这属于增强，而不是新的算法。虽然对我来讲，新加两份文件会非常容易，但那完全没必要，对于用户来讲，多几个选择，尤其是A增强了B但是完全可以做到和B一样的工作的时候（即忽略权重），这种选择很鸡肋。做加法永远比做减法容易，而当我们想要做减法的时候，难度可能会大很多。
Best wishes！
CalvinKirs
On 07/28/2020 19:18，lidong dai<da...@gmail.com> wrote：
I also think as a new algorithm will be better
Best Regards
---------------
DolphinScheduler(Incubator) PPMC
Lidong Dai 代立冬
dailidong66@gmail.com
---------------
Yichao Yang <10...@qq.com> 于2020年7月19日周日 下午6:24写道：
Hi,
I got your point. And from the point of the code logic and the functions
that can be used after code implementation, I agree that the weighted
polling algorithm is indeed applicable to scenarios without setting weights.
But in terms of module division and architecture design, the weight will
be coupled into the existing simple round robin algorithm. In fact, the
weighted round robin algorithm is also a separate algorithm, which is
different from the simple round robin algorithm in terms of
design[1].
What I want to express in my last email[2] is that from the perspective of
module division and enhancement of module clarity, Is it better
to introduce weighted rotation training and other algorithms into new
algorithms instead of adding weight logic to the previous algorithms?
--------------
我明白你的观点，从代码逻辑以及实现后能展现出来的功能角度上，我是同意加权轮询算法确实是适用于没有设置权重的场景的。
但是，在模块划分上和架构设计上，其实是把权重这部分给耦合进了现有的简单轮训算法中了。加权轮训算法其实也是一种单独的算法，和简单轮训算法在设计角度和功能上是不一样的[1]。
我在上一份邮件中表达的意思是说，从模块划分，增强模块清晰度的角度，我们把加权轮训等算法以新的算法引入，而不去在以前的算法上加入权重的逻辑是不是会更好点呢？
[1] https://blog.csdn.net/bohu83/article/details/79669051
[2]
https://lists.apache.org/thread.html/r05005385cd0a52be8e601fd1e76a8dacbf3a24c679a3a435eb1d9c8a%40%3Cdev.dolphinscheduler.apache.org%3E
If you have any question or suggetion, welcome to put forward~
Best,
Yichao Yang
------------------ 原始邮件 ------------------
发件人:
"dev"
<
acmrabbit@gmail.com>;
发送时间: 2020年7月19日(星期天) 下午5:31
收件人: "dev"<dev@dolphinscheduler.apache.org>;
主题: Re: Re： [DISCUSS]Load balancing weight
Hi, I think you have misunderstood. In fact, the weighted smooth polling
algorithm is also suitable for scenarios where no weight is set.
Best wishes！
CalvinKirs
On 2020/07/19 03:13:10, "Yichao Yang" <1048262223@qq.com> wrote:
> Hi,
>
>
> I got your point.
>
>
> I agree with your about the all algorithms you mentioned should be
measured by weight. But I think we should keep the original algorithm which
is not based on weight selection. and the algorithm based on weight should
appear in a new algorithm.
>
>
> In my opinion. Actually in our production environment, there are few
differences in the performance of the machines as you mentioned. In most
cases, the machine performance and configuration are consistent in our
production environment.
>
>
> So there are two scenarios for this weight based algorithm. The first
one is the one you mentioned. The second may occur when a user wants to use
half of the performance of machine A, rather than full use of machine A.
>
>
> So is it better to keep the original algorithm and also provide those
new selection algorithms based on weight?
>
>
>
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------ 原始邮件 ------------------
>
发件人:
"dev"
<acmrabbit@gmail.com>;
> 发送时间: 2020年7月18日(星期六) 晚上10:12
> 收件人: "dev"<dev@dolphinscheduler.apache.org>;
>
> 主题: Re: [DISCUSS]Load balancing weight
>
>
>
> Let me give an example. It might be better. Suppose my computer A has
twice
> the processing power of computer B. Then I have three tasks.
According to
> the polling algorithm, A will handle two and B will handle one. Isn't
our
> purpose of load balancing to make each machine work evenly?
>
> Best wishes！
> CalvinKirs>
>
> On 2020/07/18 13:56:38, lidong dai <d...@gmail.com> wrote:
> > I have a little confused about adding a weight to what you said,
like>
> > assign tasks to machines in turn, there is normally no
weights>
> >
> >
> >
> > Best Regards>
> > --------------->
> > DolphinScheduler(Incubator) PPMC>
> > Lidong Dai 代立冬>
> > dailidong66@gmail.com>
> > --------------->
> >
> >
> > CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
> >
> > > Hi, everyone. I’m currently designing weight-based load
balancing. My>
> > > point of view is that all load balancing needs to be done
based on
> weights,>
> > > because load actually means that machines with good
performance are>
> > > expected to handle more tasks. Therefore, weights should be
Reflected
> in>
> > > each load balancing algorithm, it should be a default
attribute in
> load>
> > > balancing.>
> > >>
> > > Hope to get your suggestions.>
> > >>
> > >>
> > > Best wishes！>
> > > CalvinKirs>
> > >>
> > >>
> >
Re: Re： [DISCUSS]Load balancing weight
Posted by lidong dai <da...@gmail.com>.
I also think as a new algorithm will be better
Best Regards
---------------
DolphinScheduler(Incubator) PPMC
Lidong Dai 代立冬
dailidong66@gmail.com
---------------
Yichao Yang <10...@qq.com> 于2020年7月19日周日 下午6:24写道：
> Hi,
>
>
> I got your point. And from the point of the code logic and the functions
> that can be used after code implementation, I agree that the weighted
> polling algorithm is indeed applicable to scenarios without setting weights.
>
>
> But in terms of module division and architecture design, the weight will
> be coupled into the existing simple round robin algorithm. In fact, the
> weighted round robin algorithm is also a separate algorithm, which is
> different from the simple round robin algorithm in terms of
> design[1].
>
>
> What I want to express in my last email[2] is that from the perspective of
> module division and enhancement of module clarity, Is it better
> to introduce weighted rotation training and other algorithms into new
> algorithms instead of adding weight logic to the previous algorithms?
>
>
>
> --------------
>
>
> 我明白你的观点，从代码逻辑以及实现后能展现出来的功能角度上，我是同意加权轮询算法确实是适用于没有设置权重的场景的。
>
>
>
> 但是，在模块划分上和架构设计上，其实是把权重这部分给耦合进了现有的简单轮训算法中了。加权轮训算法其实也是一种单独的算法，和简单轮训算法在设计角度和功能上是不一样的[1]。
>
>
>
> 我在上一份邮件中表达的意思是说，从模块划分，增强模块清晰度的角度，我们把加权轮训等算法以新的算法引入，而不去在以前的算法上加入权重的逻辑是不是会更好点呢？
>
>
> [1] https://blog.csdn.net/bohu83/article/details/79669051
> [2]
> https://lists.apache.org/thread.html/r05005385cd0a52be8e601fd1e76a8dacbf3a24c679a3a435eb1d9c8a%40%3Cdev.dolphinscheduler.apache.org%3E
>
>
> If you have any question or suggetion, welcome to put forward~
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------ 原始邮件 ------------------
> 发件人:
> "dev"
> <
> acmrabbit@gmail.com>;
> 发送时间: 2020年7月19日(星期天) 下午5:31
> 收件人: "dev"<dev@dolphinscheduler.apache.org>;
>
> 主题: Re: Re： [DISCUSS]Load balancing weight
>
>
>
> Hi, I think you have misunderstood. In fact, the weighted smooth polling
> algorithm is also suitable for scenarios where no weight is set.
>
> Best wishes！
> CalvinKirs
>
> On 2020/07/19 03:13:10, "Yichao Yang" <1048262223@qq.com> wrote:
> > Hi,
> >
> >
> > I got your point.
> >
> >
> > I agree with your about the all algorithms you mentioned should be
> measured by weight. But I think we should keep the original algorithm which
> is not based on weight selection. and the algorithm based on weight should
> appear in a new algorithm.
> >
> >
> > In my opinion. Actually in our production environment, there are few
> differences in the performance of the machines as you mentioned. In most
> cases, the machine performance and configuration are consistent in our
> production environment.
> >
> >
> > So there are two scenarios for this weight based algorithm. The first
> one is the one you mentioned. The second may occur when a user wants to use
> half of the performance of machine A, rather than full use of machine A.
> >
> >
> > So is it better to keep the original algorithm and also provide those
> new selection algorithms based on weight?
> >
> >
> >
> >
> >
> > Best,
> > Yichao Yang
> >
> >
> >
> >
> > ------------------ 原始邮件 ------------------
> >
> 发件人:
> "dev"
> <acmrabbit@gmail.com>;
> > 发送时间: 2020年7月18日(星期六) 晚上10:12
> > 收件人: "dev"<dev@dolphinscheduler.apache.org>;
> >
> > 主题: Re: [DISCUSS]Load balancing weight
> >
> >
> >
> > Let me give an example. It might be better. Suppose my computer A has
> twice
> > the processing power of computer B. Then I have three tasks.
> According to
> > the polling algorithm, A will handle two and B will handle one. Isn't
> our
> > purpose of load balancing to make each machine work evenly?
> >
> > Best wishes！
> > CalvinKirs>
> >
> > On 2020/07/18 13:56:38, lidong dai <d...@gmail.com> wrote:
> > > I have a little confused about adding a weight to what you said,
> like>
> > > assign tasks to machines in turn, there is normally no
> weights>
> > >
> > >
> > >
> > > Best Regards>
> > > --------------->
> > > DolphinScheduler(Incubator) PPMC>
> > > Lidong Dai 代立冬>
> > > dailidong66@gmail.com>
> > > --------------->
> > >
> > >
> > > CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
> > >
> > > > Hi, everyone. I’m currently designing weight-based load
> balancing. My>
> > > > point of view is that all load balancing needs to be done
> based on
> > weights,>
> > > > because load actually means that machines with good
> performance are>
> > > > expected to handle more tasks. Therefore, weights should be
> Reflected
> > in>
> > > > each load balancing algorithm, it should be a default
> attribute in
> > load>
> > > > balancing.>
> > > >>
> > > > Hope to get your suggestions.>
> > > >>
> > > >>
> > > > Best wishes！>
> > > > CalvinKirs>
> > > >>
> > > >>
> > >
Re： Re： [DISCUSS]Load balancing weight
Posted by Yichao Yang <10...@qq.com>.
Hi,
I got your point. And from the point of the code logic and the functions that can be used after code implementation, I agree that the weighted polling algorithm is indeed applicable to scenarios without setting weights.
But in terms of module division and architecture design, the weight will be coupled into the existing simple round robin algorithm. In fact, the weighted round robin algorithm is also a separate algorithm, which is different from the simple round robin algorithm in terms of design[1].
What I want to express in my last email[2] is that from the perspective of module division and enhancement of module clarity, Is it better to introduce weighted rotation training and other algorithms into new algorithms instead of adding weight logic to the previous algorithms?
--------------
我明白你的观点，从代码逻辑以及实现后能展现出来的功能角度上，我是同意加权轮询算法确实是适用于没有设置权重的场景的。
但是，在模块划分上和架构设计上，其实是把权重这部分给耦合进了现有的简单轮训算法中了。加权轮训算法其实也是一种单独的算法，和简单轮训算法在设计角度和功能上是不一样的[1]。
我在上一份邮件中表达的意思是说，从模块划分，增强模块清晰度的角度，我们把加权轮训等算法以新的算法引入，而不去在以前的算法上加入权重的逻辑是不是会更好点呢？
[1] https://blog.csdn.net/bohu83/article/details/79669051
[2] https://lists.apache.org/thread.html/r05005385cd0a52be8e601fd1e76a8dacbf3a24c679a3a435eb1d9c8a%40%3Cdev.dolphinscheduler.apache.org%3E
If you have any question or suggetion, welcome to put forward~
Best,
Yichao Yang
------------------ 原始邮件 ------------------
发件人: "dev" <acmrabbit@gmail.com>;
发送时间: 2020年7月19日(星期天) 下午5:31
收件人: "dev"<dev@dolphinscheduler.apache.org>;
主题: Re: Re： [DISCUSS]Load balancing weight
Hi, I think you have misunderstood. In fact, the weighted smooth polling algorithm is also suitable for scenarios where no weight is set.
Best wishes！
CalvinKirs
On 2020/07/19 03:13:10, "Yichao Yang" <1048262223@qq.com> wrote:
> Hi,
>
>
> I got your point.
>
>
> I agree with your about the all algorithms you mentioned should be measured by weight. But I think we should keep the original algorithm which is not based on weight selection. and the algorithm based on weight should appear in a new algorithm.
>
>
> In my opinion. Actually in our production environment, there are few differences in the performance of the machines as you mentioned. In most cases, the machine performance and configuration are consistent in our production environment.
>
>
> So there are two scenarios for this weight based algorithm. The first one is the one you mentioned. The second may occur when a user wants to use half of the performance of machine A, rather than full use of machine A.
>
>
> So is it better to keep the original algorithm and also provide those new selection algorithms based on weight?
>
>
>
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------ 原始邮件 ------------------
> 发件人: "dev" <acmrabbit@gmail.com>;
> 发送时间: 2020年7月18日(星期六) 晚上10:12
> 收件人: "dev"<dev@dolphinscheduler.apache.org>;
>
> 主题: Re: [DISCUSS]Load balancing weight
>
>
>
> Let me give an example. It might be better. Suppose my computer A has twice
> the processing power of computer B. Then I have three tasks. According to
> the polling algorithm, A will handle two and B will handle one. Isn't our
> purpose of load balancing to make each machine work evenly?
>
> Best wishes！
> CalvinKirs>
>
> On 2020/07/18 13:56:38, lidong dai <d...@gmail.com> wrote:
> > I have a little confused about adding a weight to what you said, like>
> > assign tasks to machines in turn, there is normally no weights>
> >
> >
> >
> > Best Regards>
> > --------------->
> > DolphinScheduler(Incubator) PPMC>
> > Lidong Dai 代立冬>
> > dailidong66@gmail.com>
> > --------------->
> >
> >
> > CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
> >
> > > Hi, everyone. I’m currently designing weight-based load balancing. My>
> > > point of view is that all load balancing needs to be done based on
> weights,>
> > > because load actually means that machines with good performance are>
> > > expected to handle more tasks. Therefore, weights should be Reflected
> in>
> > > each load balancing algorithm, it should be a default attribute in
> load>
> > > balancing.>
> > >>
> > > Hope to get your suggestions.>
> > >>
> > >>
> > > Best wishes！>
> > > CalvinKirs>
> > >>
> > >>
> >
Re: Re��� [DISCUSS]Load balancing weight
Posted by Kris Calvin <ac...@gmail.com>.
Hi, I think you have misunderstood. In fact, the weighted smooth polling algorithm is also suitable for scenarios where no weight is set.
Best wishes！
CalvinKirs
On 2020/07/19 03:13:10, "Yichao Yang" <10...@qq.com> wrote:
> Hi,
>
>
> I got your point.
>
>
> I agree with your about the all algorithms you mentioned should be measured by weight. But I think we should keep the original algorithm which is not based on weight selection. and the algorithm based on weight should appear in a new algorithm.
>
>
> In my opinion. Actually in our production environment, there are few differences in the performance of the machines as you mentioned. In most cases, the machine performance and configuration are consistent in our production environment.
>
>
> So there are two scenarios for this weight based algorithm. The first one is the one you mentioned. The second may occur when a user wants to use half of the performance of machine A, rather than full use of machine A.
>
>
> So is it better to keep the original algorithm and also provide those new selection algorithms based on weight?
>
>
>
>
>
> Best,
> Yichao Yang
>
>
>
>
> ------------------ 原始邮件 ------------------
> 发件人: "dev" <ac...@gmail.com>;
> 发送时间: 2020年7月18日(星期六) 晚上10:12
> 收件人: "dev"<de...@dolphinscheduler.apache.org>;
>
> 主题: Re: [DISCUSS]Load balancing weight
>
>
>
> Let me give an example. It might be better. Suppose my computer A has twice
> the processing power of computer B. Then I have three tasks. According to
> the polling algorithm, A will handle two and B will handle one. Isn't our
> purpose of load balancing to make each machine work evenly?
>
> Best wishes！
> CalvinKirs>
>
> On 2020/07/18 13:56:38, lidong dai <d....@gmail.com> wrote:
> > I have a little confused about adding a weight to what you said, like>
> > assign tasks to machines in turn, there is normally no weights>
> >
> >
> >
> > Best Regards>
> > --------------->
> > DolphinScheduler(Incubator) PPMC>
> > Lidong Dai 代立冬>
> > dailidong66@gmail.com>
> > --------------->
> >
> >
> > CalvinKirs <ac...@163.com> 于2020年7月18日周六 下午5:55写道：>
> >
> > > Hi, everyone. I’m currently designing weight-based load balancing. My>
> > > point of view is that all load balancing needs to be done based on
> weights,>
> > > because load actually means that machines with good performance are>
> > > expected to handle more tasks. Therefore, weights should be Reflected
> in>
> > > each load balancing algorithm, it should be a default attribute in
> load>
> > > balancing.>
> > >>
> > > Hope to get your suggestions.>
> > >>
> > >>
> > > Best wishes！>
> > > CalvinKirs>
> > >>
> > >>
> >