RE: RE: [SSI-users] PostgreSQL/MySQL working on top of openSSI

[Mulyadi S. <a_m...@te...>]

Hello=20

> possibly some database cleanup.  The hard part is being able to run
> multiple instances of the database on the same data at the same time.
> I would quite surprised and impresses if MySQL already had that
> capability.  But then again, without it, I'm not sure what EMIC is
> claiming.

After carefully re-read the architecture white paper of EMIC HA/Load Bala=
ncing=20
for MySQL, it does the following:
1. Interceptor and SQL Parser.=20
2. replication engine
3. Cluster Communication

"Cluster Wide IP" is used to intercept SQL query directed to MySQL instan=
ce.=20
To achieve this purpose, EMIC create some patches so MySQL can be=20
instantiated on each node using cluster wide IP and do replication

So, I do self correction here.=20

However, it seems that we can also achieve it using: Distributed Lock=20
Manager+Cluster File System (openGFS maybe)+Cluster Wide IP ..... Anyone =
has=20
comment about it?

> We have no plans to do the latter (moving just a single thread), since
> the co-ordination of the shared address space would be quite disruptive
> to the the linux MM system.  As the the ability to move an entire threa=
d

I see. Well, that came to my mind 2 years ago when I do some benchmark us=
ing=20
PVM on top of MOSIX. I drew conclusion that fine grained application (tha=
t=20
use multithreading) is best kept on same node. This is true if you use si=
mple=20
connection like Fast Ethernet. Infiniband/Myrinet might help with Remote =
Data=20
fetching (I forget the name of the feature) but still this will cost much=
 on=20
remote communication

About moving entire thread group, that sounds like feasible idea. Which=20
threading library is used as experimental environment? CMIIW I think=20
different threading library use different implemenation, such as=20
synchronization strategy etc......

> group, John Byrne is quite aware of how to do in the context of the
> current OpenSSI process migration and MM capability and now that rel 1.=
0
> is out, he may find time to implement it.

BTW, 2.6 is out. Is this mean openSSI will need major reconstruction to a=
dapt=20
new kernel? I heard that 2.6 implement different MM mechanism.......

> the data structures are set up to share.  System V shared memory object=
s
> are instantiated in Linux with inodes the shmfs and what OpenSSI has
> done is to stack the CFS filesystem on top of the shmfs so CFS provides
> the coherency, just like it does for shared mapped files.

Brilliant idea......:-) Maybe i need to test this one.....

NB: Maybe you remember that we had interview session 6 months ago, but is=
 is=20
discontinued. Do you still have time to continue the interview?

regards

Mulyadi