Benchmarks

Framework Comparison

Standard vault program (PDA derivation + system transfer CPI for deposit, direct lamport manipulation for withdraw):

Quasar matches hand-written Pinocchio (-17 CU) while providing derive macros, auto-generated clients, and constraint validation. The gap from Anchor comes from zero-copy vs Borsh deserialization, compile-time codegen vs runtime dispatch, and no_std by default.

Measuring CU

Static profiler — deterministic instruction count from the binary's .text section. Fast iteration, per-function breakdown, but counts all instructions including unreachable branches:

quasar profile

Runtime measurement — actual CU consumed by the SVM. Exact match to on-chain behavior. Use QuasarSVM or Mollusk:

let result = svm.process_instruction(&instruction, &accounts);
result.assert_success();
println!("CU: {}", result.compute_units_consumed);

Both are useful: the profiler for fast iteration and hotspot identification, runtime for precise on-chain-matching numbers.

Optimization Checklist

Inline instruction handlers — #[inline(always)] on handlers (the default in scaffolded code). Removing it can add 50-100 CU.

Minimize CPI — each cross-program invocation has fixed overhead. When you own the accounts, direct lamport manipulation is cheaper than a system program transfer.

Store bumps — find_program_address searches for the bump (expensive). If the bump is stored in the account, use bump = expr to verify it directly (cheap).

Use --expand and --watch — find the widest bars in the flamegraph, optimize them, and catch regressions as you edit:

quasar profile --watch --expand