Add model stats to Ping Pong example#265
Merged
Merged
Conversation
Terminal GVT/RT payloads now carry packets generated/finished deltas,
minimal and non-minimal finished-chunk deltas, per-rail stalled-chunk
deltas, and vc_occupancy/terminal_length snapshots. Router payloads gain
per-port stalled-chunk deltas, vc-occupancy sums, and queued-byte
snapshots. The virtual-time (analysis LP) path carries the same fields
via new dfly_cn_ross_sample / dfly_router_ross_sample structs, kept
deliberately separate from dfly_cn_sample / dfly_router_sample so the
CODES-native sampling on-disk format is untouched.
Also fixes three pre-existing bugs:
(1) terminal mstat_sz was a static 48 B, but the collect fn writes
40 + 8*num_rails, so multi-rail configs overflowed the heap; the
size is now computed at runtime like the router already does.
(2) surrogate freeze memset the terminal state and dropped the
instrumentation array pointers, so GVT collection during frozen
mode dereferenced NULL; the pointers are now carried across
freeze/unfreeze and the collect/sample fns NULL-guard, emitting
zeros for hidden network state.
(3) the terminal VT sample fn initialized its appended-array pointer
from an uninitialized buffer value (a wild write if --model-stats=3
was ever used); it now uses the array-after-struct pattern the
router fn already used.
New scalar shadows are wired into check_terminal_state; the new arrays
are listed in the save/check exclusion comments. Payload layouts are
documented in doc/model-stats-binary-format.md (later commit in this
series).
Register st_model_types for the nw-lp server LP with every mode implemented: event-type tracing, a 44-byte GVT/RT delta payload (ping/pong send/recv counts, bytes sent, RTT sum/count), and virtual-time sampling via a new svr_sample struct with forward and reverse callbacks. App-level RTT is measured by stamping the PING send time into the message and echoing it back in the PONG; the computed RTT is stored in the message so the reverse handler can subtract it exactly. Two accumulator sets are kept -- scalars for GVT/RT and svr_sample for VT -- following the dragonfly convention, so the --model-stats=4 modes do not zero each other's deltas. Registration mirrors model-net-synthetic.c and is gated on g_st_ev_trace / g_st_model_stats / g_st_use_analysis_lps before codes_mapping_setup(). Comments are deliberately didactic since this is teaching code. svr_msg grew 16 bytes, so message_size goes 440 -> 456 in all four config templates. The config headers gain a model-stats run-line note, and the tutorial comment warns that --vt-samp-end must be set explicitly because g_tw_ts_end is 24h here.
Add a byte-level spec of ross-stats-model.bin (sample_metadata + model_metadata framing, GVT/RT) and ross-stats-analysis-lps.bin (lp_metadata framing, VT), plus the per-LP payload layouts for the ping pong server and the dally terminal/router. It is written for downstream consumers (NetMaestro) so the layout no longer lives only in C code. Caveats are documented: records are not time-sorted across PEs, buffer-overflow drops can occur, the final partial interval is not emitted, RT deltas can go slightly negative under optimistic sync, and GVT and RT share an accumulator under --model-stats=4 while VT is independently complete. scripts/parse-ross-model-stats.py is a stdlib-only reference parser that emits per-LP-type CSV, dispatching payloads by size with --num-rails / --num-qos / --radix and exiting non-zero on framing errors so tests can assert parseability. codes-vis-readme.md gets a cross-link to the spec, an "nw-lp" typo fix, and its implemented-LP list brought up to date.
example-ping-pong-model-stats.sh runs GVT mode (--model-stats=1, --num-gvt=100 to keep output around 1 MB; sampling every GVT would be ~50x larger) and asserts ross-stats-model.bin is non-empty and round-trips through the reference parser with records for all three LP types and zero unknown payload sizes. example-ping-pong-model-stats-modes.sh runs --model-stats=4 and asserts the analysis-lps file exists and parses with virtual-time records present; RT record counts are wall-clock dependent, so they are not asserted. Both scripts soft-skip the parser checks when python3 is absent, since python3 is not a hard test dependency, and both are registered in the test-shell-files list.
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
Member
Author
|
FYI @marySalvi and @bnmajor |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Ping pong example will be integrated into NetMaestro soon, so this adds the model stats to be collected by ROSS instrumentation layer to the terminal LP in that example, as well as adds some additional metrics to what's collected in dragonfly dally. This also adds some documentation and a reference python parser and some tests that run if python3 is available.