Skip to content

benchmarks

Jump to bottom
github-actions[bot] edited this page Nov 7, 2026 · 1 revision

Benchmarks

This page records the reproducible performance baselines for XISOSharp and a methodology comparison against the reference tools. Results are informational: the byte-level tests, not wall-clock numbers, are the compatibility guarantee.

Hardware and runtime

Component Value
CPU AMD Ryzen 9 7900 (12C/24T, 3.70 GHz base)
RAM 32 GiB
System disk ADATA LEGEND 840 NVMe SSD
OS Windows 11, build 10.0.26300.9818
Runtime .NET 10.0.14, x64 RyuJIT (AVX-512)
BenchmarkDotNet 0.15.8

The corpus is deterministic and generated on the fly; no game data is used.

In-process library benchmarks

Command:

dotnet run --project XISOSharp.Benchmarks -c Release -- --filter "*XisoEndToEnd*" --exporters markdown
Method Mean Allocated
Pack (1 MiB tree -> ISO) 4.518 ms 12,439 KB
OpenAndList (open + full tree walk) 330.0 us 70.3 KB
Hash (CRC32/MD5/SHA-1/SHA-256/SHA3-256) 14.556 ms 1,028 KB
CisoCompress (level 6) 8.949 ms 12,969 KB
CisoDecompress 2.895 ms 1,314 KB
Extract (tree -> disk) 4.783 ms 225 KB
CciCompress (level 9, single part) 2.753 ms 2,080 KB
CciDecompress 2.761 ms 1,096 KB
LargeImageRandomRead (512 x 4 KiB reads from a 64 MiB file) 12.664 ms 1,038 KB

All values are the BenchmarkDotNet mean after at least three measured iterations. Pack/CisoCompress/CciCompress include writing the output file; Extract includes creating the destination tree on disk.

CLI comparison

Command (script in the repository):

pwsh -File scripts/bench-compare.ps1 -Iterations 2 -SizeMiB 1024

Corpus: one 1 GiB tree (8 files, deterministic pseudo-random content), measured once after a warmup. Wall clock includes process startup and cache effects.

Create (source tree -> ISO)

Tool Median Min Max MiB/s
XISOSharp -c 742.2 ms 653.3 ms 831.0 ms 1379.7
extract-xiso -c 1042.6 ms 482.3 ms 1603.0 ms 982.1
xdvdfs pack 732.4 ms 609.1 ms 855.7 ms 1398.1

Extract (ISO -> source tree)

Tool Median Min Max MiB/s
XISOSharp -x 813.3 ms 704.2 ms 922.3 ms 1259.1
extract-xiso -x 637.3 ms 600.3 ms 674.4 ms 1606.7
xdvdfs unpack 1272.6 ms 1227.5 ms 1317.6 ms 804.7

XboxKit is not in the plain-XISO create/extract tables: its -x extractor only accepts Redump-layout images with a video partition, not a standalone XISO. The BattleTests harness still compares XboxKit's Redump verbs (petrify/video/random/seed/trim/wipe/zar/rebuild) on synthetic and real Redump images.

Methodology

  • In-process BenchmarkDotNet runs isolate the library: no process startup, JIT and file-system cache controlled by BenchmarkDotNet's pilot/warmup loop.
  • CLI comparison starts a fresh process per iteration, like a user would. A framework-dependent XISOSharp CLI adds roughly 0.4 s of .NET startup on this host; extract-xiso and xdvdfs start in tens of milliseconds. At the 1 GiB corpus size the data path dominates, but the startup overhead remains visible. Use the in-process numbers for library throughput and the CLI numbers for end-to-end user-wait estimates.
  • Output files are deleted before each measured iteration. The corpus and all scratch outputs live under the temp corpus root.
  • Results depend on disk, CPU, OS cache state and power plan. Re-run the commands on the target hardware; do not compare numbers across machines as a pass/fail gate.

Reproduce

# In-process library benchmarks (all end-to-end cases)
dotnet run --project XISOSharp.Benchmarks -c Release -- --filter "*XisoEndToEnd*"

# CLI comparison (1 GiB corpus, 2 measured iterations)
pwsh -File scripts/bench-compare.ps1 -Iterations 2 -SizeMiB 1024

# Generate the public synthetic corpus used by tests/BattleTests
dotnet run --project XISOSharp.TestDataGenerator -c Release -- --corpus TestData/corpus --force

The comparison script only deletes its own %TEMP%\xiso_bench_compare workspace or the explicit -CorpusRoot passed to it.

Interpreting the numbers

  • Create is dominated by file reads and sector writes; the three tools are within roughly 1.4x on this corpus.
  • XISOSharp extraction was slower than native extract-xiso here but faster than xdvdfs unpack. Extraction includes creating the destination tree and per-file pre-sizing; --jobs 2..64 can overlap independent file copies on SSDs.
  • CCI is cheaper than CISO on this synthetic tree because its per-sector payloads are smaller; real game data with mixed entropy produces different ratios.
  • The 64 MiB random-read case is a smoke test for the image facade and block device path, not a storage benchmark.

Clone this wiki locally