Conference: Winning Strategies for Deploying Low-Latency Technologies Notes
On Monday I attended the Winning Strategies for Deploying Low-Latency Technologies conference. Unfortunately due to workload a colleague and I only managed to catch the keynote and Session 1: Latest innovations in low-latency high-performance messaging software, part of the Informatica Program – The Mission Continues – Towards Zero Latency.
I thought Ben Stephens (Electronic Trading & Quantitative Prime Brokerage Origination, Nomura International) keynote discussion was very informative, touching on many interesting data points around low latency trading:
- 10G network connection
- NYSE co-lo
- Market data processing – currently the US market is more mature in terms of FPGA hardware to process US market data, for European market data one still needs to build a bespoke solution rather than buy off the shelf
- Tech spend for a low latency solution is around $1-2m per year, which probably doesn’t cover the fraction of a cent per trade profits being made from strategies
- HFT mostly concentrate on futures, FX, equities for all three regions, with on average 100-200 employees
- Is co-lo required for every piece of the solution? co-lo is expensive. Possibly better to figure out where the biggest bang for buck is, and spend wisely from a technical solution perspective
Mention was made of two data centres in the UK:
- Equinix (Slough) – advantage of being outside London. Better for BATS and Chi-X
- Interxion. Closer to Frankfurt, more appropriate for a derivatives play?
Discussion around market data processing choices:
Then there was the execute debate. Become a member and get regulated, or act offshore?
Interesting when the audience was asked how many knew about FPGA’s, only half raised their hand. If I recall correctly, the general keynote view was that FPGA’s where becoming more main stream. Specifically the Nomura NXT platform uses FPGA’s to repeat calculation – market data processing and risk check calculations. The FPGA solution saves 10-15 microseconds over software, but offers a consistent latency for the calculations, unlike the software equivalent.
Moving to the Informatica Program – The Mission Continues – Towards Zero Latency session post the keynote, generate the following notes on the UM Streaming/UM Persistence/UM Queuing stack. 2012 should see the following features added to UM
- Inter-Thread communication model – taking advantage of the LMAX distruptor
- Zero object generation – specifically on source events
- Advanced stripping and optimized recovery
Martin Thompson, CTO of Informatica customer LMAX then discussed:
- Mechanical Sympathy – the abstraction nightmare
- Cache/Cache lines
- Benchmarks – and micro bursts
- Contention resources
- Shared nothing
- Software is not engineering, its creative and science
I did notice the sessions I attended were video’s, so hopefully the entire conference will be on-line soon.
