Trading Infrastructure - (IT)

Discussion in 'Technical Analysis' started by WallstYouth, Apr 8, 2006.

  1. I found a pretty high level overview of a trading network, I'm guessing a modern trading ticket processing infrastructure environment as shown here:

    But I'm having a bit of time breaking it all down so anyone who maintains or has designed a trading environment maybe could shed some light.

    I'm guessing the data coming in from the exchanges are raw data that then get processed by the feed handlers? What is the role of the feed handlers? and ticker feed processor

    Times ten seems to be a caching rdbms on stereoids where all transactions happen in memory, is this standard in most trading environments?? I've never heard of it till now.

    What is the role of the ticket planet output network? Is that a vendor provided network? something through reuters, comstock etc..?

    This information will be really helpfull to me thanks in advance.
  2. This is almost standard vanilla market data driven architecture. It is virtually impossible to deviate from it. The architecture assumes a direct feed model (i.e., raw data coming from exchanges, OPRA, CME, eCBOT, etc), instead an aggregating feed (i.e., Reuters, HyperFeed/Comstock, etc). The feed handlers just parses (maybe or maybe not validate the data) the incoming raw data and pass it on to the ticker processor (which performs aggregation, NBBO, etc), then push them (most likely pub/sub or multicast) to the subscribers (internal workstations), some of the server-side engines (like fair value engines, risk engines) are the "local" subscribers. The "Times-Ten" is most likely what the firm trying to sell, hence a product name, I have never heard of them, and I have been in this business for a long time. But the architecture doesn't look terribly interesting, I do something very similar in my own, where I wrote my own pub/sub mechanism. TIBCO maybe the transport mechanism, and RMDS handler are the edge handlers, and RMDS would do exactly what the Times Ten do, probably better in most cases. Comstock is a straight point-to-point feed, then internal broadcast model, so it doesn't take in raw data feeds.

    I have been watching Wombat for quite a few years now, it is a very robust architecture, but I have heard implementation costs around $10-15M minimum (which is about 2-3x a standard RMDS implementation), but this should shave a few ms from the aggregation latency.
  3. Awesome feedback!!!

    Thank you.