Verbosity and Logging Cost

Why logging gets expensive

The expensive part of logging is often string construction, not only file I/O. Calls like sprint(), deep object prints, and large sformatf payloads can dominate runtime in hot monitor and scoreboard paths.

[PERF] logging cost anatomy

  report call
    -> verbosity check
    -> message construction (sformatf/sprint)
    -> report server routing
    -> output sink (file/stdout)

  Critical insight:
    if message is built before verbosity guard,
    you pay cost even when log is suppressed.

Guard expensive formatting

function void write_act(bus_txn t);
  if (get_report_verbosity_level() >= UVM_HIGH) begin
    `uvm_info("SCB_TRACE", t.sprint(), UVM_HIGH)
  end
  compare(t);
endfunction

• Check verbosity before calling sprint() in hot functions.

• Prefer concise identifiers over full object dumps at low verbosity.

• Keep UVM_FULL for targeted replay runs only.

Targeted verbosity instead of global noise

Use component- and ID-specific verbosity controls to inspect one area without slowing the whole run.

# Global quiet mode
simv +UVM_TESTNAME=axi_long_random +UVM_VERBOSITY=UVM_LOW

# Raise only scoreboard diagnostics
simv +UVM_TESTNAME=axi_long_random +UVM_VERBOSITY=UVM_LOW \
     +uvm_set_verbosity=uvm_test_top.env.scb,SCB_TRACE,UVM_HIGH,run

# Raise one monitor path only
simv +UVM_TESTNAME=axi_long_random +UVM_VERBOSITY=UVM_LOW \
     +uvm_set_verbosity=uvm_test_top.env.axi_agt.mon,MON,UVM_HIGH,run

[REG] logging policy tiers

  farm default:
    global UVM_LOW
    no broad sprint dumps

  triage replay:
    scoped UVM_HIGH for one component ID
    preserve performance elsewhere

  deep debug:
    short run + UVM_FULL with strict time budget

Useful message design rules

• Include compact keys (id, addr, phase, status).

• Avoid redundant full transaction dumps in loops.

• Emit one summary line per N events for throughput tracking.

Report server and sink considerations

Central report processing can become a bottleneck in massive logs. Keep report IDs stable and avoid overly dynamic formatting that bloats log volume.

// Lightweight periodic stats instead of per-transaction dumps
int txn_seen = 0;

function void write_act(bus_txn t);
  txn_seen++;
  if ((txn_seen % 1000) == 0) begin
    `uvm_info("SCB_RATE", $sformatf("txn_seen=%0d", txn_seen), UVM_LOW)
  end
  compare(t);
endfunction

# Quick estimate of log volume and heavy IDs
wc -l out/logs/axi_long_random.log
rg "UVM_INFO|UVM_WARNING|UVM_ERROR" out/logs/axi_long_random.log | head -n 20

Key takeaways

• Logging overhead is mostly formatting plus report routing in hot paths.

• Guard expensive message construction behind verbosity checks.

• Prefer targeted component verbosity over global verbosity increases.

• Use compact periodic summaries to retain observability at scale.

Common pitfalls

• Calling sprint() regardless of effective verbosity.

• Running full-regression with UVM_HIGH/UVM_FULL globally enabled.

• Producing verbose logs without stable IDs for parser tooling.