Filter trace events with corrupted CUPTI durations in _compress_df#325
Open
firoz1905 wants to merge 2 commits intofacebookresearch:mainfrom
Open
Filter trace events with corrupted CUPTI durations in _compress_df#325firoz1905 wants to merge 2 commits intofacebookresearch:mainfrom
firoz1905 wants to merge 2 commits intofacebookresearch:mainfrom
Conversation
meta-cla
bot
added
the
CLA Signed
|
@firoz1905 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D98705186. |
Summary: Pull Request resolved: facebookresearch#324 OSS HTA uses ijson with the yajl C backend for memory-efficient streaming of large traces. However, yajl has a signed int64 limit and strict character validation that causes parsing failures on traces with: - Large integers (e.g. Comms Id values exceeding int64 max) - Invalid UTF-8 bytes - Unescaped quotes from old Kineto versions Internal HTA does not hit these issues because it uses Python json.loads (arbitrary precision integers) with a strip_invalid_bytes fallback. This diff adds two error handling mechanisms to OSS HTA: 1. parse_trace_dict: logs a clear error and raises on UnicodeDecodeError, telling the user the trace file is corrupted. 2. parse_trace_dataframe: wraps ijson parsing in try/except and falls back to the JSON backend when yajl fails. This preserves ijson memory efficiency for normal traces while gracefully handling traces with valid JSON that yajl cannot parse (e.g. large integers). Differential Revision: D98693551
Summary: Some GPU traces have corrupted timestamps from CUPTI, producing nonsensical event durations (e.g., 236 years). These corrupted values propagate through HTA's DataFrame and skew all downstream analysis (GPU hours, iteration times, comm/IO breakdowns). This diff adds an acceptance check in `_compress_df()`, the single chokepoint that all trace parser backends (json, ijson, ijson_batched) flow through. Events with negative or excessively large durations are dropped at parse time, before any downstream consumer sees them. **Threshold: 7 days (604,800,000,000 us)** Query over last 30 days on `ai_infra.gpu_trace_stats`: | Duration Bucket | Actual Range (ms) | Trace Count | |-----------------|------------------------------------------|-------------| | 0-1s | 0 - 1,000 | 13,404,751 | | 1s-10s | 1,000 - 10,000 | 67,420,645 | | 10s-1min | 10,000 - 60,000 | 7,364,988 | | 1min-1hr | 60,000 - 3,503,332 | 792,591 | | 1hr-1day | 4,007,737 - 53,174,381 | 5 | | 1day-7days | 201,830,019 - 286,093,774 | 2 | | 7days-1year | 610,791,801 - 29,085,041,461 | 14 | | 1year+ | 55,971,542,156 - 9,223,372,033,022 | 86 | 99.999% of traces are under 1 hour (max 3.5M ms). Everything above 7 days is clearly corrupted (max is 9.2e15 ms = INT64_MAX/1000, classic CUPTI overflow). Related: D96421400 (downstream fix in Durin post-processing) Differential Revision: D98705186
firoz1905
force-pushed
the
export-D98705186
branch
from
8e38189 to
2e3559d
Compare
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.
Summary:
Some GPU traces have corrupted timestamps from CUPTI, producing nonsensical
event durations (e.g., 236 years). These corrupted values propagate through
HTA's DataFrame and skew all downstream analysis (GPU hours, iteration times,
comm/IO breakdowns).
This diff adds an acceptance check in
_compress_df(), the single chokepointthat all trace parser backends (json, ijson, ijson_batched) flow through.
Events with negative or excessively large durations are dropped at parse time,
before any downstream consumer sees them.
Threshold: 7 days (604,800,000,000 us)
Query over last 30 days on
ai_infra.gpu_trace_stats:99.999% of traces are under 1 hour (max 3.5M ms). Everything above 7 days
is clearly corrupted (max is 9.2e15 ms = INT64_MAX/1000, classic CUPTI overflow).
Related: D96421400 (downstream fix in Durin post-processing)
Differential Revision: D98705186