- CME gives you co-location facilities for $6,000 per connection (not counting co-location facility charges), so that you can place your servers closer to the exchange and reduce latency.
- Eurex's co-location gets you a round-trip of below 10 milliseconds, when compared with 29 milliseconds for a connection in London and 128 milliseconds for a connection from Chicago.
- ICE is worrying about a new notion of quality of service: no more than 0.1% of the messages should get processed in worse than 50 milliseconds, once a new trading engine is in place.
- Acworth says: In the old days, a trader might be watching four screens at the same time, commented one speaker. Now its two rows of four, with one row displaying the markets and the other row displaying network conditions.
- People like the fact that in C++ you get to control when garbage collection is done, so as to avoid unpredictable glitches in performance. (Ugh!)
- The average order size on the E-mini has dropped to 2 contracts.
- The biggest systems are nearing a million messages a second.
- Some firms have started moving functions from software to FPGA in the quest for speed.
A million messages a second is a lot! I remember in the late 1990s, when I worked on the PRISM system (which does the realtime VaR at NSE), our goal was to be able to handle 100 trades/s, which is 200 VaR calculations a second. We were thrilled because the system evolved into exceeding 1000 trades a second.
For those interested in algorithmic trading, you may like to see Chapter 5 of the MIFC report.