High-performance trading architecture.

Discussion in 'Hardware' started by Waterloo, Sep 27, 2005.

  1. Hey All

    I've tried searching both Google and ET with little success. Here's what I'm hoping for...

    I would like to find some resources (books/papers etc..) that provide an outline/architecture of a high performance trading environment. For example, if I were a hedge fund, what type of hardware+software architecture would I have? ie. how many data feeds(main + failover), raid systems, backups, db's(Oracle/SQL/in memory). Would I use a message queues? How about clusters(metooxx)?

    I hope you get the idea. Would you really use TS and a T1 if you had 100mil fund? I wouldn't think so, so where would you start to design an "enterprise solution"? Are there off the self solutions?

    In my head I see this (I'm pulling the ranges out of thin air) ...

    Joe DayTrader -> HF w/ 10mill uses TS and a T1 with a Raid.
    HF w/ 10 mil+ -> HF w/200 mill uses ??????(fill in with resources here!)
    HF w/200+ builds their own.

    Any resources would be appreciated.
  2. You are overthinking it in my opinion. There are no standars here becaues Internet is getting faster more quickly than traders know what to do with it. You can trade a decent fund like about 300 Million from your bedroom with your underwear on and a cable modem...........how do I know that :)
  3. I agree, but I'm thinking if you are a mechanical based trader, ie. you've got custom code, you would probably want that in version control sytems - cvs, sourcesafe.

    If you do backtesting and own your own data, how do you back up the data. Since you own it, it's an asset that you want to protect from loss.

    I guess if I had my money at Citadel I would feel secure that a computer failure wouldn't put them out of business for a week. Plus their systems would still know that I'm a client after the failure because my records were not lost. That's why I'm asking about the architecture. I'm sure you can do the bedroom and cable thing but it's risky.

    I'm thinking more professional trading corporation architecture, rather that at-home day trader.
  4. ozzy


    First make a few million than come back and ask this question.

  5. Maybe you're right, then I could just pay some one else to figure it out. :)

    The engineer in me just craves the knowledge though.
  6. ozzy


    I'm the same way. When you have the money and time it shouldn't be too hard to figure out what you need.

    Good Luck
  7. A couple of things. First of all, a $200M fund is comparatively small these days. I was part of a $200M fund, and we had < 10 people (and only 2 technologists). Keep in mind that a $200M fund returning 15% only has $8M annual fee (at 1 and 20).

    The standard answer will be, it depends on the trading strategy. if you are a $1B fund focusing on distress securities or say fundamental sector based, the type of technology is not terribly high-end, you can get away with Bloomberg stations and one or two custom applications. I was surprised when I talked to cto of one of the largest fund in the world, they are using standard Dells running Bloomberg, no TIB, a little SQL server for historical, that's it. But then again, they are highly sector and fundamental research based.

    Now, if you are talking about high-frequency quant fund of $1B+, that's a completely different story. However, these funds are usually *not* exchange members, so taking the raw feed from the exchanges (say NYSE / CME) is not an option for them. Therefore you are back to the standard Reuters / Wombat setup, which, while a bit pricy ($2M, 2x DS3s, 5-10 servers to start), also get you the TIBCO licenses to do messaging. Easy, that's how most of these funds do it anyways. But a $1M annual license fee is not that steep in relation to a $1B fund returning 15% (that's annual fee of $35M). Now, some funds hate TIB (your usual variety of storm, managability, etc, etc), and write their own messaging backbone, and that's no small task (and expensive).

    Funds that require high-end compute farms are even more rare, and these tend to be built for special purposes. The main cost in these setups are not the hardware any more (a 64 node Egenera will only run $2M or so), it is the specialized knowledge to maintain and support them, and these cost a lot (it is not un-usual for these people to cost $200-250 annually). Heck, a high-end data center cost > $5M each.

    Also, I am not sure why you are talking about raid, etc, by definition, all servers follow a spec, from the vendor, very very standard and basic stuff.

    Just my $0.02,


  8. Babak


    MAESTRO might be able to help you out (his work station rocks my world)

    btw if you're reading this check your pm big M. I'm like 2 blocks away and can't wait to get in on a tour.
  9. rufus, awesome answer, thanks.

  10. MAESTRO? What, isn't he in Toronto! I can barely afford gas in Colorado, let alone pay Petro Canada prices!! Though, maybe I can scrap together enough Canadian Tire money. :)

    The GST kills me every time we go back to visit. We were in Belleville (yeah, I know) over the summer and my wife buys something at Walmart. Pulls out $8 bucks since the item was $7. It was 8.05! It's like Martin was sucking the loonie right out of pocket.

    #10     Sep 27, 2005