Hello!

I found a very informative page on scaling the Image Server

http://www.clearcanvas.ca/dnn/Portals/0/ClearCanvasFiles/Documentation/UsersGuide/ImageServer/2_0/index.html?hardware_configuration.htm

Is there a similar guide for scaling the RIS server?

In case I needed redundancy,or need to share load on different servers.

I did some investigation, and found a very nice tool called HAProxy, which can share the load on multiple servers: http://haproxy.1wt.eu/

Also, I found that web farms using NHibernate use a distributed 2nd level cache, primarily memcached.

So, I would like to play with the following scenario: two servers running RIS connecting to single DB server,  another server running memcached with shared cache, and finally a server running HAProxy to divide the load; HAProxy would have

a VIP that clients connect to.

Can anyone comment on any other issues that might arise with this configuration?

Do I need to configure HAProxy for sticky sessions?

One other option I was thinking of was to turn off second level caching in NHibernate, and forgetting about memcached. But that depends on the read/write ratio I guess.

If ratio is low, then might make sense.

Interesting problem.....

Any insight into this situation would be very greatly appreciated!

 

Cheers,

Igor

 

 

 

 

 

 

 

 

Also, as per MSDN, I will try reducing the lease timeout on the RIS WCF services, to say 15 seconds, to avoid client affinity to a particular server, which would defeat load balancing.

Hi,

There is no guide to scaling RIS server at this time, but the approach you've described seems generally sound. A couple of things to consider:
1. You can try disabling the 2nd level cache, but I suspect that performance will suffer. Memcached is probably a good option, even you don't have a dedicated machine for it.
2. You shouldn't need sticky sessions - each client/server request is atomic. I don't think you need to worry about lease timeout either, because the client is establishing a new wcf channel for every call to the server.
3. Shreds other than the application server were not designed to support running multiple instances. Therefore you will need to ensure that you disable these shreds on all but one of your servers. This can be accomplished by ExtensionSettings in the app config, or by physically removing the plugin ClearCanvas.Ris.Shreds.dll.

If you get this setup going, please let us know about your experiences.

Thank very much for the tips. I suppose the ImageServer needs sticky sessions because of how DICOM connections work, where a client negotiates an association with a particular server, and must maintain its association with that server until the association is closed.

For the lease timeout, I believe that a new wcf channel may reuse an existing connection from the connection pool, so channel != socket necessarily. See this for example:

http://kennyw.com/work/indigo/173

As for the shreds other than the application server shreds, would those be the image availability and publication shreds?


Thanks again,
Igor

Yes, the image availability and publication shreds.

Thanks!

One other issue just occurred to me: synchronizing time stamps. I assume the RIS server
stores the time when patients are registered, reports are created, etc. If I have multiple RIS servers, would I need to have a central time provider, to keep time stamps consistent?

Thanks!

Igor

Yes, the easiest thing to do would be to configure the OS to keep your server clocks synchronized.