gazagnaire.org
Logging, runtime tracing, and the obs performance profiler for OCaml
observe: GC-section CPU time agrees with the budget cpu
The GC section showed CPU time = wall x domains (capacity), which read
34.229s right under the budget cpu=2.498s and made GC overhead (% of CPU)
meaningless for an idle guest (0.02% of wall vs 0.22% of on-CPU). Use the
budget cpu (wall - blocked - gc) so the two numbers match and the overhead
is a fraction of time actually spent computing.
author
Thomas Gazagnaire
date
(Jun 25, 2026, 9:20 PM -0700)
commit
f03ebce2
f03ebce21673030de7e94bb801bb07280dd5f54d
parent
0cc6a74e
0cc6a74eaf6d305ef506b7b8521ca28af3064600
+4
-2
+4
-2
lib/obs/memtrace_eio.ml
+4
-2
lib/obs/memtrace_eio.ml
···
1125
1125
let running =
1126
1126
Hashtbl.fold (fun _ (f : fiber_acc) acc -> acc +. f.run_time) t.fibers 0.
1127
1127
in
1128
-
(* CPU time approximated as wall per domain summed; cap GC by it. *)
1129
-
let cpu_time = wall_time *. float_of_int (Stdlib.max 1 n_domains) in
1128
+
(* On-CPU time: wall minus what was blocked off-CPU or in GC -- the same cpu
1129
+
the budget reports, so the GC section agrees with it. An idle unikernel is
1130
+
mostly waiting, not on-CPU, so this is far below wall. *)
1131
+
let cpu_time = Stdlib.max 0. (wall_time -. t.blocked_time -. gc_time) in
1130
1132
let gc_overhead = safe_div gc_time cpu_time in
1131
1133
let blocked_frac = safe_div t.blocked_time wall_time in
1132
1134
let gc_frac = safe_div gc_time wall_time in