Profiling and Base Performance Optimization in `iex2h5`#

Overview#

This article documents the profiling journey of the iex2h5 application, focusing on identifying and optimizing base performance bottlenecks using Valgrind Callgrind and KCachegrind.

Tools Used#

perf
valgrind --tool=callgrind
kcachegrind
Flat profile, call graph, and callee map analysis

Key Hotspots Identified#

1. `std::chrono::operator<=>`#

auto std::chrono::operator<=><long, std::ratio<1l, 1000000000l>, long, std::ratio<1l, 1000000000l> >

Inclusive: ~2.9B cycles
Self: ~2.9B cycles
Calls: ~57M

Appears in:

if (now > today + this->start) ...
if (now > today + this->stop)  ...
if (now - last_time >= this->heart_beat_interval) ...

2. `std::chrono::duration::count()`#

std::chrono::duration<long, std::nano>::count() const

Self: ~1.8B cycles
Calls: ~231M
Pure accessor, but called from every comparison

3. `iex::transport_t::transport_handler(...)`#

Inclusive: ~56B cycles
Self: ~3B cycles
Calls: ~19M

This is the driver of the event loop, and indirectly responsible for millions of time comparisons and filter applications.

4. `std::unordered_map` Internals#

std::_Hashtable, _Hash_node, and std::__detail symbols total:
Inclusive: ~8.3B cycles
High lookup and insert volume due to symbol mapping

Performance Analysis#

Hot paths show excessive use of std::chrono::duration and related operators.
Cost comes from abstraction layering: templates, type casting, virtual calls, and inline barriers.

Optimization Strategies#

Replace Time Comparisons#

Original:

if (now > today + this->start) ...

Optimized:

int64_t now_ns = now.time_since_epoch().count();
int64_t start_ns = today.time_since_epoch().count() + this->start.count();
if (now_ns > start_ns) ...

Use Raw Timestamps in Hot Loops#

Avoid instantiating std::chrono::time_point inside per-tick message processing.

Flatten Container Usage#

Replace std::unordered_map<K,V> with std::vector<uint64_t> + binary search where keyspace is dense and flat.

Lessons Learned#

Observation	Recommendation
High cost of `std::chrono` ops	Use `int64_t` timestamps
Operator overloads not optimized	Avoid templated comparisons in hot path
Containers impact performance	Consider flat maps for known contract ID range
Filters and EMA computation	Profiled but not yet optimized — low hanging fruit later

Conclusion#

This profiling session helped uncover unexpected hotspots in high-level constructs. By peeling away abstraction where it matters (timing logic, container lookups), substantial CPU time can be reclaimed. Profiling is not just about finding bugs — it's about finding what matters.

Authored by Steven Varga, June 2025

Profiling and Base Performance Optimization in iex2h5#