[ic] Interchange and clustering: LVS for http(s), and MySQL replication for read-only clustering?

[Mike Heins]

Quoting Dan Browning (danb@cyclonecomputers.com):
> Friends, Romans, countrymen, lend me your ears:
> 
> How does IC interact with clustering/load balancing?  Specifically, I'm
> looking at implementing something like the following:
> 
> Web Clustering
> --------------
> HTTP (and HTTPS/SSL) clustering, Software based, using LVS
> (www.linuxvirtualserver.org).  I plan to use mod_ssl on each IC server to
> decrypt the SSL traffic (there should be a fair amount).
> 
> This is the most likely scenario because it is open source.  LVS supposedly
> works mostly on the layer-2 (and sometimes 3,4) levels.  As I understand IC,
> it would require somehow centrally storing all the IC-related directories,
> such as (for example) '/var/lib/interchange' and '/usr/local/interchange'.
>  - What is the best way to have that kind of smart, central storage?
> 	- Is NFS on Linux stable enough?  v3?
> 	- How much read/write activity does IC have during normal browse-and-buy
> useage?
> 	- What X to Y ratio of NFS server beef to web server would be required?
> 		- For example:  a 10-disk RAID-10 ultra160 array on a Quad-Zeon server
> acting as NFS  ---->  10 P-3 700Mhz 512MB web (IC) servers?
>  - Would another way be better than NFS?
> 	- Coda, is it reliable enough?
> 	- GFS?
> 	- InterMezzo?

I can't answer all these issues with authority. I see no reason why NFS
wouldn't work -- you could also store sessions in SQL. NFS on Solaris
is reputed to have the best and most reliable locking, but I hear later
versions of Linux NFS are OK. I was never able to break Linux NFS locking
in the tests I did with Interchange split over multiple machines, but I
can't say those were extensive and really stressful.

>  - I imagine that I could just install interchange on the NFS server, and
> map the '/usr/local/interchange' and '/var/lib/interchange' to each web
> server (client) as read-write.
> 	- But will I run into locking and contention issues?

Not other than database issues and possibly the UserTrack log (but you can
turn that off in favor of cookie-based tracking). Any other logs that are
frequently written could be a contention issue, but the logData routine
in Interchange cycles the lock for every write chunk.

If you use DBM at all, you don't want it to be a read-write database. Most
IC tables are not open for write very often, only on order.

> 	- Does IC hold any files open for a write for long periods of time?

No. In the normal file-based sessions, each session is an individual
file, so there are no "hotspots" that I know of.

One of these days we will write a white paper on this issue, I hope.

-- 
Akopia, Inc., 131 Willow Lane, Floor 2, Oxford, OH  45056
phone +1.513.523.7621 fax 7501 <heins@akopia.com>

Friends don't let friends use Outlook. -- Bob Blaylock