TimeBase codecs performance

In this JMH microbenchmark we measure peformance of timebase message codecs.

Test description

We measure speed of encoders and decodes separately for three kind of messages:

1. Empty message. JSON equivalent of this message is

{
  "$type": "deltix.qsrv.hf.pub.InstrumentMessage",
  "symbol": "ETHUSDT",
  "timestamp": "2025-04-02T13:56:00.337Z"
}

note

This empty message is practically a "noop" operatoin for codes since symbol/timestamp/message length are system-level fields and appear in message header (outside of payload codecs).

2. Incremetnal market data update containing single Trade inside the payload package. JSON equivalent:

{
  "$type": "deltix.timebase.api.messages.universal.PackageHeader",
  "symbol": "ETHUSDT",
  "timestamp": "2025-04-02T13:56:00.337Z",
  "originalTimestamp": "2025-04-02T13:56:00.257Z",
  "receiveTimestamp": "2025-04-02T13:56:00.336Z",
  "packageType": "INCREMENTAL_UPDATE",
  "entries": [
    {
      "$type": "deltix.timebase.api.messages.universal.TradeEntry",
      "exchangeId": "BINANCE",
      "matchId": "2293043024",
      "price": "1881.52",
      "side": "SELL",
      "size": "0.003"
    }
  ],
  "receivedTime": "2025-04-02T13:56:00.336Z"
}

3. Full MBP order book snapshot (20 bids + 20 offer). JSON equivalent:

{
  "$type": "deltix.timebase.api.messages.universal.PackageHeader",
  "symbol": "BTCUSDT",
  "timestamp": "2025-04-02T13:56:00.300Z",
  "receiveTimestamp": "2025-04-02T13:56:00.247Z",
  "packageType": "VENDOR_SNAPSHOT",
  "entries": [
    {
      "$type": "deltix.timebase.api.messages.universal.L2EntryNew",
      "exchangeId": "BINANCE",
      "price": "85664.11",
      "size": "0.01974",
      "level": 0,
      "side": "BID"
    },
    {
      "$type": "deltix.timebase.api.messages.universal.L2EntryNew",
      "exchangeId": "BINANCE",
      "price": "85664.1",
      "size": "0.00185",
      "level": 1,
      "side": "BID"
    },
    ... 18 more entires ...
  ],
  "firstSequenceNumber": 0,
  "lastSequenceNumber": 0,
  "receivedTime": "2025-04-02T13:56:00.247Z"
}

The source code for this benchmark is available in:
deltix.qsrv.hf.pub.codec.perf2.Benchmark_Codecs1

Environment

Results were measured on developer workstation:

• CPU: i7-13850HX, 2100 Mhz, 20 Cores
• RAM: 32G
• SSD: Intel 660p NVMe
• OS: Windows 11 PRO
• OpenJDK 17.0.12
• TimeBase 5.6.161

Results

Benchmark                 (codecType)                            (msgFile)  Mode  Cnt     Score     Error  Units
Benchmark_Codecs1.decode     compiled  test_message_1_vendor_snapshot.json  avgt   15  1390.452 ±  37.552  ns/op
Benchmark_Codecs1.decode     compiled       test_message_2_inc_update.json  avgt   15   131.405 ±   1.907  ns/op
Benchmark_Codecs1.decode     compiled          test_message_3_minimal.json  avgt   15     2.258 ±   0.068  ns/op
Benchmark_Codecs1.decode  interpreted  test_message_1_vendor_snapshot.json  avgt   15  6650.045 ± 363.036  ns/op
Benchmark_Codecs1.decode  interpreted       test_message_2_inc_update.json  avgt   15   375.875 ±   6.072  ns/op
Benchmark_Codecs1.decode  interpreted          test_message_3_minimal.json  avgt   15     4.174 ±   0.099  ns/op
Benchmark_Codecs1.encode     compiled  test_message_1_vendor_snapshot.json  avgt   15  1335.759 ±  35.540  ns/op
Benchmark_Codecs1.encode     compiled       test_message_2_inc_update.json  avgt   15    58.077 ±   1.200  ns/op
Benchmark_Codecs1.encode     compiled          test_message_3_minimal.json  avgt   15     0.795 ±   0.035  ns/op
Benchmark_Codecs1.encode  interpreted  test_message_1_vendor_snapshot.json  avgt   15  5853.044 ±  97.254  ns/op
Benchmark_Codecs1.encode  interpreted       test_message_2_inc_update.json  avgt   15   284.026 ±   7.198  ns/op
Benchmark_Codecs1.encode  interpreted          test_message_3_minimal.json  avgt   15     2.137 ±   0.045  ns/op

Analysis

The benchmark highlights a consistent pattern: compiled codecs are an order of magnitude faster than interpreted codecs across all tested message types.

• Minimal message (NOOP codec) The smallest possible message shows the lowest latency:

  • Compiled: ~2.3 ns/op (decode), ~0.8 ns/op (encode)
  • Interpreted: ~4.2 ns/op (decode), ~2.1 ns/op (encode)

  Given that this kind of message is practically empty, we are essentially measuring encoding framework overhead.

• Incremental update (single Trade)

  • Compiled: ~131 ns/op (decode), ~58 ns/op (encode)
  • Interpreted: ~376 ns/op (decode), ~284 ns/op (encode)
    This shows a 3–5× advantage for compiled codecs, with latencies in the sub-100 ns range for encode.

• Full order book snapshot (40 entries)

  • Compiled: ~1390 ns/op (decode), ~1336 ns/op (encode)
  • Interpreted: ~6650 ns/op (decode), ~5853 ns/op (encode)
    For large payloads, compiled codecs are still ~5× faster. Interpreted latencies approach 6–7 μs per operation.

Overall, the results demonstrate linear scaling of compiled codecs with message size, while interpreted codecs accumulate overhead with each field.

Conclusion

The fastest result is the minimal message encode with compiled codecs at ~0.8 ns/op.
For practical workloads, compiled codecs remain consistently 3–7× faster than interpreted, making them the clear choice for high-throughput scenarios.