A highly subjective
and emotive subject. For me its Java.
A C++ expert that is
not also an expert in Java is probably because they don’t believe in Java or
have just played with it for a year or two. They simply don’t have the
experience to compare fairly. I know because that used to be me !
"just because you don’t know how, doesn’t mean
its not possible"
A Java program can
be 100 times faster than a functionally similar C++ program … fact.
The key is not the
language, C, C++ and Java all end up as machine code. The biggest factor in
performance is the application design and in particular its threading model.
Note many patterns
for low latency (core affinity, pooling etc) work just as well in Java as they
do in C++
I have been asked by
many why bother with Java why not just use C++ ? For the same reason we don’t use assembler
anymore. The compilers are highly optimised and use of Java will minimise
production downtime and maximise productivity while reducing project costs and
avoid pain of trying to hire highly specialised and expensive expert devs /
hardware experts.
Productivity is much
higher with Java, it has much better tooling than C++. Java systems excel with ease of maintenance,
productive graduates, simpler code.
The argument against
Java always comes back to Jitter from GC. To be honest that’s simple to avoid
with object pooling. The really painful grief comes from the JIT compiler. Its
really hard to write good warmup code that exercises enough of the code base to
get best optimised code.
On C++ and avoiding
virtual methods. I have seen low latency C++ systems which avoid any use of
virtual methods to avoid vtable lookup. A lot of the code ends up looking like
C at which point I wonder why bother. Object Orientation is key in trading systems
to avoid code bloat, maximise code reuse and significantly reduce development
time while increasing quality. One nice feature in Java is the ability at
runtime to determine that a method is not extended and for virtual methods to
be inlined. Ofcause if you are messing about with dynamic class generation or
custom class loaders you may well end up with recompilation jitter (honestly
don’t do it -> KISS)
KISS - thanks to
David Straker for the best IT lesson I ever had. Back in 1988 in my first
developer role, I wrote a really complex C installation program for HPWord. Its
was damn complex and I thought it was clever, David kicked it back into touch
and told me 'Keep It Simple Stupid'. Many years on and I can honestly say
that’s the most important principle I learnt in software development. If
someone cant understand the code they cant maintain it.
In my experience
java systems are built in half the time, cost a third as much and are much more
stable in production. That said, ultra low latency requires object reuse which
in itself brings risk regardless of language. I developed SubMicroTrading on a
dual core Intel 1.6Ghz, 4GB RAM,
11" Dell Adamo …. On that I could run exchange simulator, client
simulator and the complete SMT Trading System !
It wouldn’t of been possible with C++. The dev cycle, code + compile +
run + rinse repeat would have been factorially slower.
On other languages,
SCALA is a GC sink and while functional code is great for prototyping
strategies its implicit nature generates a lot of pretty inefficient code. The
argument that the same code in SCALA compiles to the save code in Java misses
the point. In my experience the algo written in SCALA can be much more
efficently rewritten in Java. SCALA is not suitable for ultra low latency in my
opinion.
On GO, I havent used
it but was asked recently about using it, honestly my reply of "why"
… whats it going to give me that I don’t have with Java with regard to ultra
low latency. If I didn’t have over 4 years of IP in SubMicroTrading I would consider
it …. But it’s a risk, maybe its tomorrows C# ?
FPGA Hybrid Systems
… personally over the years I have seen various claims and promises which
didn’t materialise (details of which I cant divulge due to the NDA's banks
signed). Beware the smoke and mirrors, understand the true cost of latency in
the hybrid system with wire to wire independent metrics for key use cases. Also
ensure you understand the edge risks, eg hidden latency, partial TCP stacks,
FPGA to main memory latency, and for the strategy atomic book reading and
skipping intermediate book updates.
Please do not poke
me about C++ or FPGA, I quite frankly am tired of all the arguing and hype. If
you want to use C++ / C / FPGA or whatever then that’s your choice I hope you
enjoy.
