Explore the results

One message. Every operation.

Choose a FIX message and compare encoding, checked parsing and parsing + extraction together for every language. Each table is ranked by measured p50 within its language and operation.

Campaign: HP · Intel Xeon Gold 6254 · Linux 5.14.0-687.44.1.el9_8.x86_64 · measurement CPUs 16–17; launcher CPUs 0–7 · 18 October 2026 · measured LibHFT source 5917330b7c2c. Peer versions and pinned inputs are listed under method and evidence.

FIX message

NewOrderSingle · 25 measured adapters · 73 operation results. Lower is better.

Compare p50 and p99 at a glance. Expand CDF & all percentiles for the top-five-engine chart and the full ranked table. Missing results are labelled, never plotted as zero.

See the NewOrderSingle message and all three tests →

JavaScript is disabled: NewOrderSingle results for all three operations and every language remain available below, including expandable charts and full percentile tables. Download the CSV for other message types.

Percentiles are rounded batch-average nanoseconds per operation: each of the 50,000 samples averages 32 in-process operations, so they are not individual-call or network latencies. CDFs connect measured percentile anchors; they are reconstructed profiles, not raw-sample empirical CDFs. What each test measures →

Same-host codec comparison

C++ and C

Native FIX engines and codec libraries, including LibHFT C++, QuickFIX/C++, NexusFIX, Fix8, hffix, Fixpp, llfix and libtrading. llfix encoding is ranked here; its receive cells include SDK timestamp, sequence and callback work and are listed separately.

Encoding

Field values → FIX bytes

NewOrderSingle · encoding · ns/operation
Engine / codec	p50	p99
LibHFT C++	49	50
hffix	127	132
llfix	321	328
NexusFIX	363	380
libtrading	445	457
QuickFIX/C++	2,614	2,624
Fix8	3,065	3,088
Fixpp	3,088	3,145

CDF & all percentiles

Reconstructed CDF of the fastest 5 qualifying engines by p50 — Percentile-anchor reconstruction, not a raw-sample empirical CDF. Up to five engines with matching evidence; the table retains the full roster. Open CDF ↗

NewOrderSingle · encoding · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT C++	49	50	50	51
2	hffix	127	130	132	133
3	llfix	321	324	328	330
4	NexusFIX	363	371	380	388
5	libtrading	445	451	457	461
6	QuickFIX/C++	2,614	2,620	2,624	2,627
7	Fix8	3,065	3,076	3,088	3,097
8	Fixpp	3,088	3,110	3,145	3,172

How this test works →

Checked parsing

FIX bytes → checked field values

NewOrderSingle · checked parsing · ns/operation
Engine / codec	p50	p99
LibHFT C++	356	358
hffix	389	397
NexusFIX	849	856
Fixpp	919	922
libtrading	1,479	1,484
QuickFIX/C++	2,095	2,103
Fix8	5,519	5,538
llfix	Receive includes SDK timestamp, sequence and callback work · listed separately, unranked

CDF & all percentiles

NewOrderSingle · checked parsing · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT C++	356	358	358	359
2	hffix	389	394	397	397
3	NexusFIX	849	854	856	858
4	Fixpp	919	920	922	924
5	libtrading	1,479	1,482	1,484	1,486
6	QuickFIX/C++	2,095	2,099	2,103	2,105
7	Fix8	5,519	5,530	5,538	5,546
—	llfix	Receive includes SDK timestamp, sequence and callback work · listed separately, unranked

How this test works →

Parsing + extraction

FIX bytes → extracted business fields

NewOrderSingle · parsing + extraction · ns/operation
Engine / codec	p50	p99
LibHFT C++	356	358
hffix	387	395
NexusFIX	852	861
Fixpp	918	921
libtrading	1,479	1,484
QuickFIX/C++	2,093	2,100
Fix8	5,511	5,529
llfix	Receive includes SDK timestamp, sequence and callback work · listed separately, unranked

CDF & all percentiles

NewOrderSingle · parsing + extraction · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT C++	356	357	358	359
2	hffix	387	392	395	397
3	NexusFIX	852	857	861	863
4	Fixpp	918	920	921	923
5	libtrading	1,479	1,482	1,484	1,486
6	QuickFIX/C++	2,093	2,097	2,100	2,102
7	Fix8	5,511	5,521	5,529	5,536
—	llfix	Receive includes SDK timestamp, sequence and callback work · listed separately, unranked

How this test works →

Same-host codec comparison

Java

Java Pure and Java/JNI beside QuickFIX/J, Philadelphia, Artio and Falcon. JNI timing includes one native codec call per message.

Encoding

Field values → FIX bytes

NewOrderSingle · encoding · ns/operation
Engine / codec	p50	p99
Philadelphia	308	1,564
LibHFT Java Pure	347	627
LibHFT Java/JNI	392	1,014
Falcon	887	1,199
Artio	1,169	2,269
QuickFIX/J	1,346	5,008

CDF & all percentiles

NewOrderSingle · encoding · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	Philadelphia	308	311	1,564	1,741
2	LibHFT Java Pure	347	535	627	2,625
3	LibHFT Java/JNI	392	497	1,014	1,201
4	Falcon	887	920	1,199	4,965
5	Artio	1,169	1,272	2,269	23,957
6	QuickFIX/J	1,346	1,375	5,008	5,656

How this test works →

Checked parsing

FIX bytes → checked field values

NewOrderSingle · checked parsing · ns/operation
Engine / codec	p50	p99
LibHFT Java Pure	585	1,179
LibHFT Java/JNI	710	1,353
Falcon	743	2,766
Philadelphia	979	2,104
QuickFIX/J	1,629	4,265
Artio	1,995	3,869

CDF & all percentiles

NewOrderSingle · checked parsing · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT Java Pure	585	587	1,179	1,259
2	LibHFT Java/JNI	710	715	1,353	1,410
3	Falcon	743	747	2,766	4,904
4	Philadelphia	979	985	2,104	5,239
5	QuickFIX/J	1,629	1,663	4,265	7,834
6	Artio	1,995	2,536	3,869	384,320

How this test works →

Parsing + extraction

FIX bytes → extracted business fields

NewOrderSingle · parsing + extraction · ns/operation
Engine / codec	p50	p99
LibHFT Java Pure	583	593
LibHFT Java/JNI	710	722
Falcon	743	767
Philadelphia	981	1,025
QuickFIX/J	1,630	1,722
Artio	1,996	2,081

CDF & all percentiles

NewOrderSingle · parsing + extraction · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT Java Pure	583	585	593	649
2	LibHFT Java/JNI	710	713	722	781
3	Falcon	743	746	767	914
4	Philadelphia	981	986	1,025	1,240
5	QuickFIX/J	1,630	1,639	1,722	1,998
6	Artio	1,996	2,010	2,081	2,410

How this test works →

Same-host codec comparison

.NET

.NET Pure and .NET Native beside QuickFIX/n and the public FIX Antenna .NET Core codec. The native flavor includes its interop boundary.

Encoding

Field values → FIX bytes

NewOrderSingle · encoding · ns/operation
Engine / codec	p50	p99
LibHFT .NET Pure	191	985
LibHFT .NET Native	265	539
FIX Antenna .NET Core	2,583	10,503
QuickFIX/n	4,782	18,952

CDF & all percentiles

Reconstructed CDF of the fastest 4 qualifying engines by p50 — Percentile-anchor reconstruction, not a raw-sample empirical CDF. Up to five engines with matching evidence; the table retains the full roster. Open CDF ↗

NewOrderSingle · encoding · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT .NET Pure	191	962	985	1,011
2	LibHFT .NET Native	265	526	539	549
3	FIX Antenna .NET Core	2,583	2,631	10,503	14,034
4	QuickFIX/n	4,782	4,811	18,952	22,481

How this test works →

Checked parsing

FIX bytes → checked field values

NewOrderSingle · checked parsing · ns/operation
Engine / codec	p50	p99
LibHFT .NET Native	501	1,812
LibHFT .NET Pure	1,004	2,662
FIX Antenna .NET Core	1,651	3,161
QuickFIX/n	4,987	7,334

CDF & all percentiles

NewOrderSingle · checked parsing · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT .NET Native	501	503	1,812	1,932
2	LibHFT .NET Pure	1,004	1,007	2,662	2,864
3	FIX Antenna .NET Core	1,651	1,670	3,161	4,416
4	QuickFIX/n	4,987	5,023	7,334	7,756

How this test works →

Parsing + extraction

FIX bytes → extracted business fields

NewOrderSingle · parsing + extraction · ns/operation
Engine / codec	p50	p99
LibHFT .NET Native	484	486
LibHFT .NET Pure	1,021	1,029
FIX Antenna .NET Core	1,648	1,690
QuickFIX/n	4,981	7,291

CDF & all percentiles

NewOrderSingle · parsing + extraction · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT .NET Native	484	485	486	487
2	LibHFT .NET Pure	1,021	1,025	1,029	1,035
3	FIX Antenna .NET Core	1,648	1,665	1,690	4,264
4	QuickFIX/n	4,981	5,015	7,291	7,676

How this test works →

Same-host codec comparison

Rust

LibHFT Rust, FerrumFIX, DFX, fixer-rs, TrueFix and NanoFIX under the same message fixtures and measured operations.

Encoding

Field values → FIX bytes

NewOrderSingle · encoding · ns/operation
Engine / codec	p50	p99
LibHFT Rust	144	145
NanoFIX	168	171
FerrumFIX	1,271	1,311
TrueFix	1,634	1,648
fixer-rs	3,811	3,826
DFX	5,815	5,862

CDF & all percentiles

NewOrderSingle · encoding · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT Rust	144	144	145	146
2	NanoFIX	168	170	171	172
3	FerrumFIX	1,271	1,290	1,311	1,321
4	TrueFix	1,634	1,644	1,648	1,651
5	fixer-rs	3,811	3,819	3,826	3,839
6	DFX	5,815	5,840	5,862	6,009

How this test works →

Checked parsing

FIX bytes → checked field values

NewOrderSingle · checked parsing · ns/operation
Engine / codec	p50	p99
LibHFT Rust	527	539
NanoFIX	1,248	1,256
TrueFix	2,036	2,042
DFX	2,538	2,547
fixer-rs	4,674	4,686
FerrumFIX	7,422	7,450

CDF & all percentiles

NewOrderSingle · checked parsing · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT Rust	527	533	539	544
2	NanoFIX	1,248	1,252	1,256	1,260
3	TrueFix	2,036	2,039	2,042	2,044
4	DFX	2,538	2,544	2,547	2,549
5	fixer-rs	4,674	4,680	4,686	4,689
6	FerrumFIX	7,422	7,438	7,450	7,461

How this test works →

Parsing + extraction

FIX bytes → extracted business fields

NewOrderSingle · parsing + extraction · ns/operation
Engine / codec	p50	p99
LibHFT Rust	528	544
NanoFIX	1,230	1,248
TrueFix	2,036	2,043
DFX	2,656	2,674
fixer-rs	4,611	4,624
FerrumFIX	7,425	7,460

CDF & all percentiles

NewOrderSingle · parsing + extraction · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	LibHFT Rust	528	537	544	550
2	NanoFIX	1,230	1,239	1,248	1,256
3	TrueFix	2,036	2,040	2,043	2,045
4	DFX	2,656	2,668	2,674	2,678
5	fixer-rs	4,611	4,619	4,624	4,627
6	FerrumFIX	7,425	7,443	7,460	7,470

How this test works →

Same-host codec comparison

Go

QuickFIX/Go is retained in the wider codec roster. There is no LibHFT Go flavor and no same-language competitor ranking to infer from this single row.

Encoding

Field values → FIX bytes

NewOrderSingle · encoding · ns/operation
Engine / codec	p50	p99
QuickFIX/Go	8,128	182,531

CDF & all percentiles

Reconstructed CDF of the fastest 1 qualifying engines by p50 — Percentile-anchor reconstruction, not a raw-sample empirical CDF. Up to five engines with matching evidence; the table retains the full roster. Open CDF ↗

NewOrderSingle · encoding · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	QuickFIX/Go	8,128	13,063	182,531	498,344

How this test works →

Checked parsing

FIX bytes → checked field values

NewOrderSingle · checked parsing · ns/operation
Engine / codec	p50	p99
QuickFIX/Go	2,405	4,637

CDF & all percentiles

NewOrderSingle · checked parsing · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	QuickFIX/Go	2,405	4,436	4,637	84,929

How this test works →

Parsing + extraction

FIX bytes → extracted business fields

NewOrderSingle · parsing + extraction · ns/operation
Engine / codec	p50	p99
QuickFIX/Go	2,401	4,650

CDF & all percentiles

NewOrderSingle · parsing + extraction · ns/operation · ranked by p50
Rank	Engine / codec	p50 (ns)	p90 (ns)	p99 (ns)	p99.9 (ns)
1	QuickFIX/Go	2,401	4,436	4,650	59,625

How this test works →

Listed separately · C++

llfix receive results, unranked.

The llfix 1.0.8 receive path is measured through the upstream LLFIX_BENCHMARK hook, without SDK patches or a substitute parser. That hook enters FixClient::process_rx_buffer, which reads the monotonic clock, updates the sequence store and dispatches the callback for every message. Those costs are inside each sample, so these six cells stay outside the codec rankings. SDK receive validations and serialization are disabled for this lane. Its three encoder cells, which include the SDK-generated SendingTime, are ranked in the C++ tables above.

Swipe the table horizontally for all percentiles.

llfix receive · HP · 18 October 2026 · rounded batch-average ns/operation
Message · operation	p50	p90	p99	p99.9
NewOrderSingle · Checked parsing	1,469	1,477	1,481	1,485
NewOrderSingle · Parsing + extraction	1,474	1,485	1,489	1,495
ExecutionReport · Checked parsing	1,741	1,747	1,753	1,758
ExecutionReport · Parsing + extraction	1,756	1,761	1,766	1,769
MarketDataSnapshot · depth 10 · Checked parsing	4,957	4,964	4,968	4,972
MarketDataSnapshot · depth 10 · Parsing + extraction	9,034	9,103	9,131	9,154

An earlier single-engine llfix run on SCAN (3 October 2026, source 211372a6) remains historical evidence with its own summary CSV, field probes, host qualification, source and checksum receipt, measurement clock, CPU configuration and snapshot. It is not merged with this campaign.

What the test measures

Three operations. Shared messages.

WRITE

Encoding

Start with field values, populate a message and serialize it through the public builder. Framing and checksum work are included only where that adapter’s build path performs them.

PARSE

Checked parsing

Parse and check the complete business-field set for D/8. For W, check symbol, entry count and best bid/offer. Numeric conversion and the typed value hash are inside timing.

READ

Parsing + extraction

Parse, read the specified business fields and fold them into a checked sink. Depth-10 market data includes all 20 ordered bid/offer entries.

Fixtures are NewOrderSingle (35=D), ExecutionReport (35=8) and a 10-level MarketDataSnapshot (35=W). The same pinned fixtures and sample counts are used throughout.

See the actual FIX messages and a worked explanation of all nine tests →

Codec cost is not deployed FIX latency.

The ranked cells exclude networking, Logon, sequence recovery, storage and application callbacks. llfix’s receive cells include its SDK callback and are listed separately, unranked. There is no offered messages-per-second target in this in-process test, and the inverse of a codec timing is not sustained session capacity. Use the full-engine comparison for that question.

From bytes to measured work

What happens in each test?

The corrected converted-values-v2 contract requires complete checked field work for both D/8 parse labels. They intentionally perform the same operation; neither is a framing-only shortcut. W distinguishes top-of-book checks from full-book extraction. Build starts from values, not by parsing or echoing the input message.

Swipe diagrams sideways on a small screen, or open a diagram to zoom.

Checked parse: receive bytes, verify the required fields.

Parse-only operation — Pinned bytes enter the public parser. D and 8 check the complete converted business-field set; W checks top of book. The typed value hash is checked inside each operation. One clock sample covers 32 repeated operations. Open diagram ↗

Parse and extract: the reads are inside the sample.

Parse-and-extract operation — Parse the pinned message, convert and hash the full common field set. D checks nine fields; 8 checks twelve; W checks all 20 entries in order. For D/8 the two parse labels now perform the same checked work. One timed sample batches 32 operations. Open diagram ↗

Build: start from values, not a received message.

Message-building operation — Start from field values, populate the public builder, serialize and consume output length. Framing depends on the adapter. One timed sample batches 32 complete build operations. Open diagram ↗

In the examples, | represents the real FIX separator SOH (ASCII 0x01). Line breaks are for readability only. Restore SOH and remove those line breaks to recover the pinned input bytes; BodyLength (9) and CheckSum (10) are calculated for those bytes.

Tags 9001 and 9002 are benchmark-specific field-work hashes, not standard FIX fields or cryptographic signatures. The oracle uses type-marked converted values: unsigned integers, scale-five fixed-point prices and ASCII strings, each followed by byte 1. The fixed July timestamps belong to the fixtures, not the October measurement date. Download the pinned examples and their source hashes.

NewOrderSingle · 35=D

A day limit order to buy 100,000 AUD/NZD at 1.08245. 11 is the client order ID, 55 the symbol, 54=1 buy, 38 quantity, 40=2 limit, 44 price and 59=0 day time-in-force.

Pinned input for both parsing tests

8=FIX.4.2|9=161|35=D|49=CLIENT|56=SERVER|34=8|52=20260703-19:28:00.123|
11=ORD123456|21=1|55=AUD/NZD|54=1|60=20260703-19:28:00.123|38=100000|40=2|44=1.08245|59=0|9002=2863910631|
10=021|

1. Checked parse parse_only: Feed the complete order into the public parser and read all nine fields: 11, 21, 55, 54, 60, 38, 40, 44, 59. Convert numeric values, fold typed values into FNV-1a and compare with 9002=2863910631. No order is submitted and no session state changes.
2. Parse and extract parse_extract: Perform the same complete checked operation as D parse_only. Both names remain for table compatibility, not as different amounts of field work. Text and numeric mutations independently verify each operation's reads.
3. Build build_D: Populate an order with the business values above through the public builder, serialize its fields and consume the output length. Sequence numbers or generated client order IDs may change between iterations. This measures constructing an order, not transmitting it or obtaining an execution.

Sample build payload · header and framing omitted

35=D|
11=ORD123456|21=1|55=AUD/NZD|54=1|60=20260703-19:28:00.123|38=100000|40=2|44=1.08245|59=0|

An illustrative payload assembled from the fixture’s business values, not a captured byte-for-byte result from every builder.

ExecutionReport · 35=8

An acknowledgement that the order is new, with all 100,000 units still open: 150=0 is a new execution event, 39=0 new order status, 151=100000 remaining quantity, 14=0 filled quantity and 6=0.00000 average fill price. This is not a fill report.

Pinned input for both parsing tests

8=FIX.4.2|9=193|35=8|49=CLIENT|56=SERVER|34=9|52=20260703-19:28:00.123|
37=EX123456|17=EXRPT1|11=ORD123456|150=0|39=0|55=AUD/NZD|54=1|38=100000|151=100000|14=0|6=0.00000|60=20260703-19:28:00.123|9002=564350835|
10=162|

1. Checked parse parse_only: Parse and perform the same complete twelve-field checked operation described below. It does not update an order book or position.
2. Parse and extract parse_extract: Parse, read and hash twelve values: order ID (37), execution ID (17), client order ID (11), execution type (150), order status (39), symbol (55), side (54), order quantity (38), remaining quantity (151), filled quantity (14), average price (6) and transaction time (60). The typed hash includes converted quantities and fixed-point price. Compare it with 9002=564350835; mismatches count as errors.
3. Build build_8: Set those twelve business fields and the message header, serialize through the public builder and retain the output length. The receive-side 9002 audit marker is not part of this build payload. Execution IDs, sequence numbers and field order may vary by adapter.

Sample build payload · header and framing omitted

35=8|
37=EX123456|17=EXRPT1|11=ORD123456|150=0|39=0|55=AUD/NZD|54=1|38=100000|151=100000|14=0|6=0.00000|60=20260703-19:28:00.123|

The report’s business fields without the receive-side audit marker; not a captured wire result from every builder.

Market data snapshot · 35=W

A ten-level AUD/NZD book: ten bids and ten offers, each sized at 1,000,000. 268=20 counts entries, not levels. Each entry contains side (269=0 bid or 269=1 offer), price (270) and size (271). Best bid/offer are 1.08241 / 1.08245.

Complete pinned input · all twenty entries, with one bid/offer pair per display line

8=FIX.4.2|9=723|35=W|49=CLIENT|56=SERVER|34=10|52=20260703-19:28:00.123|
262=REQ-AUDNZD-1|55=AUD/NZD|268=20|
269=0|270=1.08241|271=1000000|269=1|270=1.08245|271=1000000|
269=0|270=1.08240|271=1000000|269=1|270=1.08246|271=1000000|
269=0|270=1.08239|271=1000000|269=1|270=1.08247|271=1000000|
269=0|270=1.08238|271=1000000|269=1|270=1.08248|271=1000000|
269=0|270=1.08237|271=1000000|269=1|270=1.08249|271=1000000|
269=0|270=1.08236|271=1000000|269=1|270=1.08250|271=1000000|
269=0|270=1.08235|271=1000000|269=1|270=1.08251|271=1000000|
269=0|270=1.08234|271=1000000|269=1|270=1.08252|271=1000000|
269=0|270=1.08233|271=1000000|269=1|270=1.08253|271=1000000|
269=0|270=1.08232|271=1000000|269=1|270=1.08254|271=1000000|
9001=2376139772|9002=2366019821|10=214|

1. Parse only parse_only: Give the complete depth-ten message to the parser. Read symbol (55), entry count (268) and the first two prices (270), then compare their typed converted-value hash with 9001=2376139772. This is a top-of-book check, not extraction of all twenty entries.
2. Parse and extract parse_extract: Parse, read request ID (262), symbol and entry count, then visit all twenty entries in wire order. Read each entry’s side, price and size: 60 group values plus the three leading values. Convert and hash them, then compare with 9002=2366019821. Reading only the first bid/offer pair cannot satisfy this check.
3. Build build_W: Set request ID and symbol, create 268=20 and write the ten bid/offer pairs, then serialize and retain the output length. The payload contains the full book, not a two-entry shortcut. Receive-side audit tags 9001 and 9002 are omitted. Falcon W is unsupported: upstream Falcon has no market-data message model, so these three cells are not measured for it.

Sample build payload · all twenty entries, header and framing omitted

35=W|
262=REQ-AUDNZD-1|55=AUD/NZD|268=20|
269=0|270=1.08241|271=1000000|269=1|270=1.08245|271=1000000|
269=0|270=1.08240|271=1000000|269=1|270=1.08246|271=1000000|
269=0|270=1.08239|271=1000000|269=1|270=1.08247|271=1000000|
269=0|270=1.08238|271=1000000|269=1|270=1.08248|271=1000000|
269=0|270=1.08237|271=1000000|269=1|270=1.08249|271=1000000|
269=0|270=1.08236|271=1000000|269=1|270=1.08250|271=1000000|
269=0|270=1.08235|271=1000000|269=1|270=1.08251|271=1000000|
269=0|270=1.08234|271=1000000|269=1|270=1.08252|271=1000000|
269=0|270=1.08233|271=1000000|269=1|270=1.08253|271=1000000|
269=0|270=1.08232|271=1000000|269=1|270=1.08254|271=1000000|

The complete book payload without receive-side audit tags; not a captured wire result from every builder.

Read build timings at the adapter’s actual boundary.

Builders need not emit identical bytes: IDs, field order, FIX version and number formatting can differ. Most paths finish a framed FIX message; the pinned Java Pure path times header/body serialization without final 8/9/10 framing. That row is not a measurement of a completed on-wire message.

How the separate mutation checks prove field work

These are admission checks, not extra latency rankings. The harness mutates a field while keeping the message’s length and FIX checksum valid, but leaves the expected value hash stale. An adapter passes the probe only when it reports the hash mismatch.

Report extraction: change 37=EX123456 to 37=FX123356. The 9002 check must detect the changed order ID.
Full-book extraction: change the second level’s prices from 1.08240 / 1.08246 to 1.08250 / 1.08236. The best bid/offer stay unchanged; the full 9002 check must notice the deeper entries.
Top-of-book check: change 55=AUD/NZD to 55=BUD/NYD. The selected W parse_only operation must detect the stale 9001 value; rejection by a different operation cannot satisfy the probe.
Numeric conversion: change the order price from 44=1.08245 to 44=1.08254, or the last book size from 271=1000000 to 271=0100000. These preserve framing but change the converted value and must fail the selected operation’s hash check.

All 332 isolated text/numeric probes pass in the 25-adapter HP campaign. Each supported D/8/W parsing operation must reject its own mutation with a positive error count and exit code 3; rejection by another operation cannot satisfy it. Falcon has eight D/8 probes and llfix twelve receive-field probes; LibHFT C++ adds probes for its cursor and legacy entrypoints, which are diagnostic and not displayed. This is field-work evidence, not full FIX conformance or network performance.

Comparison boundaries

A missing cell is not a measurement.

The page shows each adapter’s default entrypoint for D, 8 and depth-10 W: encoding, checked parsing and parsing + extraction. Cursor, legacy and other diagnostic variants are excluded from these tables.

Falcon W is unsupported. Falcon has D and 8 results only; selecting market data shows it as not supported, without timings or a rank.
llfix receive is unranked. Its six receive cells include SDK timestamp, sequence and callback work, so they are listed separately. Its three encoder cells are ranked.
libtrading depth-10 W uses the retained 128-field capacity patch.
FerrumFIX uses the pinned v0.7.0 shim.
LibHFT Java Pure build timing covers header and body serialization without final 8/9/10 framing.

Other non-admitted, unsupported or licence-excluded targets are not shown. Their absence here is not evidence of slower performance. JMH runs, llfix reference transport tests and full-session results are separate evidence, outside this codec table.

The page shows 216 ranked D/8/W codec cells and six unranked llfix receive cells. The campaign’s summary CSV retains all 560 operation rows, including administrative messages and the C++ cursor/legacy variants, without mixing those paths into the default ranking.

Recorded methodology

One host. Frozen inputs. Audited windows.

Host	HP · Intel Xeon Gold 6254 · Linux 5.14.0-687.44.1.el9_8.x86_64 · measurement CPUs 16–17; launcher CPUs 0–7. Hardware epoch record
Campaign date	18 October 2026. This is a pinned codec snapshot, not an assertion about every later release or every deployment.
Measured source	LibHFT `5917330b7c2c7d67d01eea58379eca00bd5389bb`; later publishing revisions do not change it
Method	`converted-values-v2`: 50,000 timed samples after 5,000 warm-up samples per operation; 32 operations per clock sample
Admission	25/25 build/run gates, 332/332 extraction mutation probes, 560 operation rows and a complete clock contract
Host qualification	Passing 120-second quiet-host preflight before each of four audited segments; five-second all-PID monitor; no competing samples at or above 5% CPU in any selected adapter window
Selection	Four segments with the same source, prepared runtimes, fixtures, CPU affinity and sample counts. Contaminated windows were excluded, not averaged into the ranking.

The mutation audit changes wire-valid fields while leaving their expected field hash stale. A passing probe rejects the mismatch, demonstrating that the adapter reads the required fields rather than merely accepting a fixed message.

Peer versions and pinned inputs

Versions are shown where the campaign evidence records one. Other peers are identified by the SHA-256 of the pinned source tree or jar that was compiled and measured; the prepared manifest lists every input.

Measured versions and pinned inputs · HP · 18 October 2026
Adapter	Recorded version	Pinned input · SHA-256 prefix
LibHFT .NET Native	LibHFT source 5917330b7c2c	Measured source above
LibHFT .NET Pure	LibHFT source 5917330b7c2c	Measured source above
LibHFT C++	LibHFT source 5917330b7c2c	Measured source above
LibHFT Java Pure	LibHFT source 5917330b7c2c	Measured source above
LibHFT Java/JNI	LibHFT source 5917330b7c2c	Measured source above
LibHFT Rust	LibHFT source 5917330b7c2c	Measured source above
Artio	0.178	`artio-codecs-0.178.jar` · `d4eff79c0958`
DFX	Not version-labelled in the evidence	`dfx` · `a20c4e0933f8`
Falcon	Not version-labelled in the evidence	`falcon-0.1.0-master-SNAPSHOT.jar` · `fc2085cee1a2`
FerrumFIX	v0.7.0 shim	`ferrumfix` · `b381f8e4273a`
FIX Antenna .NET Core	Not version-labelled in the evidence	`fixantenna_net` · `c4c4ad326f9f`
Fix8	Not version-labelled in the evidence	`fix8/include` · `2c895eb218f9`
fixer-rs	Not version-labelled in the evidence	`fixer_rs` · `2cde38c9dc45`
Fixpp	Not version-labelled in the evidence	`fixpp/include` · `448c4879c106`
hffix	Not version-labelled in the evidence	`hffix/include` · `398621473830`
libtrading	Not version-labelled in the evidence	`libtrading/include` · `175aca29e54d`
llfix	1.0.8	`llfix/include` · `8b81571731fc`
NanoFIX	Not version-labelled in the evidence	`nanofix` · `c7b7e48898be`
NexusFIX	Not version-labelled in the evidence	`nexusfix/include` · `896323421dac`
Philadelphia	2.0.0	`philadelphia-core-2.0.0.jar` · `396136037b0f`
QuickFIX/C++	Not version-labelled in the evidence	`quickfix_cpp/include` · `a1a5402cc254`
QuickFIX/Go	Not version-labelled in the evidence	`quickfix_go` · `f7cc9f36cefa`
QuickFIX/J	2.3.1	`quickfixj-core-2.3.1.jar` · `d063713127ae`
QuickFIX/n	Not version-labelled in the evidence	`quickfix_n` · `91bd60612128`
TrueFix	Not version-labelled in the evidence	`truefix` · `3c64f5be84a2`

Download measured operation CSVAll 560 recorded rows; source for the tables above. Read the admission packageCounts, scope, disclosures and the digest of every campaign file. Read the admission recordHost audit replay, anchors and what the recovered package does not claim. Browse every compact resultThe 560 HP operation results with their CDFs and provenance.

Summary CSV SHA-256: dfda98a2cd6d288c44bda4666966ea4941611d2d9f59f78db01576c1a314366c. Prepared manifest SHA-256: 62b830d8b3d7ea45af627c6bb1257b2a92a103e11eec58d6b2c6aeb20692f1ac. Host qualification SHA-256: f63aae7a540c9ec65c7e9c2d41bcf9cda45cbda1c4bb75b1e6e710d629096422. Compact result set SHA-256: ace4e77ba182a372d90e086e287b96deaec65ba78ced16dbc90c647df2148ec4. Prepared manifest · machine-readable snapshot.

Reproduce the pinned measurement procedure

Use the measured source revision, prepared-manifest digest and peer pins above. Preparation and timing are separate; reserve the host CPU slot and review process activity before publishing a result.

python3 bench/cpp/fixbench/tools/open_source_fix_engines/open_source_fix_engine_stage2.py \
  --workspace "$LIBHFT_BUILD_ROOT/stage2-workspace" \
  --run-prepared --parity-audit --run-cpus 16,17 \
  --iterations 50000 --warmup 5000

This command requires the previously prepared, hash-verified workspace; it is not a standalone installation command. Another host joins this comparison only after its own qualification with matching inputs.

Historical campaign

The earlier SCAN campaign (Intel Xeon E5-2698 v4, isolated CPUs 18–19, 2–3 October 2026) measured the same 25 adapters with LibHFT source 4375a8fe and different peer pins: 320 probes and 554 operation rows. It is retained as historical evidence and is not merged into these rankings.

SCAN microbenchmark reportEvery operation, build result and mutation probe from that campaign. SCAN operation CSVAll 554 rows of the historical campaign. SCAN campaign and host reviewSource pins, repairs, qualification and preservation receipts. Microbenchmark contractFixture, API, field-work and admission boundaries.