LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

How about using both IPVS and TCPHA

To: lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Subject: How about using both IPVS and TCPHA
From: dragonfly@xxxxxxxxxxxx
Date: Sat, 19 Feb 2005 23:02:05 +0800
hi,
   The idea came from cheaney Chen,details see below.I think it is
cool.My idea is:Modify TCPHA to merge tcpha_fe with tcpha_be.That
means every node can do both dispatching(TCP Handoff) and serving
requests. In the cluster,FE run IPVS,BE run TCPHA.FE do
content-blind switch.BE do content-aware switch.If BE scheduled itself,
then serve request itself.or else it migrate the TCP connection to the
scheduled BE.By this mechanism,overload of layer-7 switch and TCP
Handoff are distributed in BEs,the advantages of layer-7 switch such as
high cache hit rate and saving server resources can be achieved.Also the
responses bypass FE.
   I am not very sure it is whether be valuable.So Your comments and
suggestions are welcome.

regards
                                                            Feb 19,2005

----- 转发来自cheaney Chen <cheaney.chen@xxxxxxxxx>的邮件 -----
   日期: Sun, 9 Jan 2005 11:09:50 +0800
  发件人:cheaney Chen <cheaney.chen@xxxxxxxxx>
回复至:cheaney Chen <cheaney.chen@xxxxxxxxx>
    主题:Re: Some Question about TCPHA
   收件人:"dragonfly@xxxxxxxxxxxx" <dragonfly@xxxxxxxxxxxx>

Hi,

  I've done the test. TCPHA runs well in my blade serer. I will use
some testing tools , like Smartbit , to measure it's performance.

Actually I am trying to develop a distributed content-aware
architecture. Front-end is a content-blind load balancer, like LVS/DR.
Content-aware function is performed by each of back-end server.

For some requests, it's good to forward it to another back-end server. 
For some requests, it's good to serve it locally.

My main concern is "High cache hit rate lead to high capacity". So the
extra forwarding overhead could benefit the system, because of less
disk I/O overhead.

There are a lots of feature need to be added to TCAPHA.

I want to modify TCPHA as follows,
1, modify TCPHA and install TCPHA_FE and TCPHA_BE in one box. Then
every node in the cluster can forward requests to other node, and get
some request from other nodes.

2. add "redirect to local " fuction in TCPHA_FE.

3. add my dispatching algorithm into TCPHA. 

4. more..... 

Could you give some advise ?! or some point needs to be concern ?

Thanks for reading this mail. And thanks for help.

Best regards,
Cheaney


----- 转发的邮件结束 -----



<Prev in Thread] Current Thread [Next in Thread>
  • How about using both IPVS and TCPHA, dragonfly <=