Serge Aleynikov serge@REDACTED
Sat Jan 7 18:07:15 CET 2006


In this particular system we are not using yaws, but do use mnesia (a 
couple of disc_copies master nodes and several ram_copies replicated in 
geograpically distributed sites).

The clients are custom host processes written in C that control Excel 
switches.  There are over 200 such clients/switches clustered in 
multiple sites.  The protocol between clients and back-end Erlang system 
is ASCII based (similar to http headers) over persistent TCP 
connections.  The Erlang system consists of three layers:

1. Proxy (load balances transactions accross back-end servers)
2. Back-end servers joined by process groups.  These also run mnesia 
ram_copies replicas.
3. Master mnesia database nodes (two nodes at two sites).

The solution also has redundancy at OSI layers: 1,2,3,4,5.

In the lab we were able to achieve the thoughput of 5000 
transactions/second using a single proxy and multiple back-end servers. 
  So in real life we'll be using multiple sites more for network 
redundancy reasons rather than performance, as the current throughput 
requirement is under 600 transactions/sec.

I would say that the most complicated part of this project was not the 
design and development in Erlang, but getting all teams involved in QA, 
deployment, support, network provisioning, server support, etc. 
coordinated, so that the project can be successfully deployed.  Yet, 
this is fairly common for large organizations...

Ernie Makris wrote:
> Hi,
> I'm curious, are you using yaws? What are the front ends that deliver
> all those transactions?
> Also, are you using mnesia?
> I'm trying to figure out the best way to use mnesia and yaws in those
> types of scenarios.
> Thanks
> Ernie
> Danie Schutte wrote:
>>Hi Serge,
>>  I can give you some references from our clients.  We only use Erlang, and 
>>it has been running in production for more than 3 years.  
>>Our clients are A grade banks and also some private companies that operate in 
>>the financial sector.
>>Another example is the article from British Telecom, the reliability of the 
>>Ericsson switches, and the volumes of calls - which also translates to 
>>The CIO can even contact me, as we answer the same questions for other banks.  
>>In actual fact we start implementations in 5 new countries during January and 
>>February, so there is a reliable track record that Erlang can deliver.
>>The one production system, which has been a pilot project for 12 months now :) 
>>- is handling 650 000 financial transactions a month, total value being 
>>managed by the system - about 30 000 000 USD per month.
>>Kind regards

More information about the erlang-questions mailing list