发一篇purescale的资料

wolfop

先总结一下
1. share disk的数据库的始创者是IBM，DB2 Z在1997年就通过sysplex实现了share everything.
2. purescale来自于DB2 FOR Z上面的架构参考
3. purescale采用了集中式的CF机制（这也是DB2 FOR Z的机制）来防止rac那种分布式锁维护的高开销
4. purescale采用了infiniband rdma技术直接远程内存访问，防止了RAC那种通过UDP的低效率的通信方式
5. purescale在处理全局页缓存上提供了比RAC更好的锁定和存储方式。
6. purescle通过这三种方式得到了更好的横向扩展性。
不过由于需要infiniband支持，估计大家没法在家里自己试验purescale了，不如RAC好试验。

wing hong

thanks for sharing. looks very interesting . we have implemented 10g RAC for a critical bank application, but got lots of complain due to various issue.

just question:

say, for a  5 member cluster , does this mean in this purescale config, there will be one CF primary server, one secondary CF server, and 3 member server  ?

How about in a 2 member cluster ?

or is CF just a whole bunch of processes, that it can coexist with member server ?

so in a 2 member cluster scenario ?

one will be CF primary as well as member server, and another will be CF secondary as well as member server.

All the global lock information is located in CF primary.

The idea of RAC using distributed lock manager is supposed to be more scalable and more robust. you can have all but one nodes down.

In theory , using centralized lock manager should be more easily to program and therefore more efficient for the code.

It is shame that we can't do purescale test at home by ourself. or is it really not any way ?

because could be very interesting to test it .  ( i have tested single instance of DB2 8.1/9.0 on PC )

wolfop

First, I don't think you can test purescale on you PC. Because it need infiniband support and current only support on P550 and P595 hardware. According to IBM plan, they will port it to all IBM's POWER sever in future. Maybe It will support normal PC running Linux with infiniband in future, maybe DB2 version 11?

For 5 member cluster, according to me understand, should means 5 DB2 instance with at least 2 CF in 7 LPAR/physical machine. Therefore, even for 2 member cluster, you need at least 4 machine or 4 LPAR to deal with it. And according to IBM's recommendation, 4 machine is better due to not sharing hardware with DB2 instance and CF.
CF is just a whole bunch of processes, but strongly recommend use separated machine/LPAR.  
For RAC's distributed lock manager, I still think it is a theory better solution, but due to IPC/software bug/network latency, always cause more problem.

Since you mentioned critical bank application, why not just use DB2 ON Z. Sysplex for Z is much mature than purescale.
IBM DB2 prueScale scale.pdf (4.75 MB, 下载次数: 128)

2010-2-27 16:08 上传

点击文件名下载附件

[ 本帖最后由 wolfop 于 2010-2-27 16:08 编辑 ]

Pythagoras

The number of CFs is NOT counted in the number of data sharing members.
Regardless of the number of data sharing members, you need at least one CF.
For high availability consideration, 2 or more CFs is highly recommended.

cngd1

应该都是主机上才能用得上的，估计没机会接触到，帮顶

wing hong

thanks guys for the clarification.

The project used Oracle RAC and Tuxedo ,MQ, and Websphere . For the app server side, use Veritas cluster solution. So in theory , this is the first step to migrate from IBM mainframe to distributed HA solution.

So supposed to more modern , and more affordable , and more time to market.

But end up with lots of performance and availability issue, so the business now is considering to use IBM mainframe again . :-)

wolfop

Seems mainframe always take the most critical application in finance, and hard to be replaced. And I never think mainframe is not modern, most newest technology are coming from mainframe.

[ 本帖最后由 wolfop 于 2010-2-28 19:19 编辑 ]

wing hong

sorry modern may not be the right word. it is just most of the time people think of mainframe application as legacy application.

anyhow a couple of projects canceled or decided to use DB2 on Mainframe rather than Oracle RAC.

So for Oracle still some way to go .