Topic · Quant Development & Systems
← All topicsSystems & Performance
33 articles · 6 checkpoints · 20 deeper reads · 7 reference notes
A standalone topic: it is on no roadmap, so read it on its own terms.
Every article, in reading order
plant a flag as you finish eachRead these first
Latency is how long one thing takes; throughput is how many things get done per second, the two can move in opposite directions, and a system tuned for one often gets worse at the other.
A CPU keeps several small, progressively faster pools of memory between itself and main RAM, and code that touches memory in a predictable, local pattern can run many times faster than code that jumps around, even doing identical arithmetic.
Modern CPUs guess which way an if-statement will go before they know for certain, and start executing that guess early, a wrong guess is thrown away at a real, measurable cost, which is why predictable branches run faster than random ones even with identical logic.
Kernel bypass lets a trading application talk to the network card directly, skipping the operating system's networking stack, to shave microseconds off every packet.
Lock-free data structures let multiple threads share data without ever blocking each other with a mutex, using hardware atomic instructions instead. They trade a simpler mental model for brutal correctness bugs, which is why a trading system reaches for them only on the handful of paths where a lock's worst case is unacceptable.
Hardware performance counters are registers built into the CPU that count real events, cache misses, branch mispredictions, instructions executed, letting you find the actual cause of slowness instead of guessing.
Then the rest