buckit/testing/singletrip-hdd

Single-Trip HDD Bench Rig

This directory is the checked-in source of truth for the ad hoc HDD benchmark rig used for phase-1 single-trip GET evaluation.

It replaces the temporary host-only copies that were previously stored under /root/singletrip-bench/ on 192.184.90.116.

Files

• start-buckit.sh
  • starts Buckit with a supplied list of local data paths
  • enables or disables the fast path with FAST_GET=0|1
• bench-warp-2m.sh
  • runs the saturated warp get 2 MiB comparison used for the earlier Docker and host pilots
• cold-curl-ttfb.sh
  • runs the host-local cold TTFB test over a list of presigned URLs
  • drops page cache before each GET
  • prints one curl TTFB value per request

Expected Host Layout

Default host workspace:

• root dir: /root/singletrip-bench
• Buckit binary: /root/singletrip-bench/bin/buckit
• Warp binary: /root/singletrip-bench/bin/warp

You can copy this directory to the host as-is and then copy or build the binaries into bin/.

Start Buckit

FAST_GET=1:

FAST_GET=1 ./start-buckit.sh /mnt/data01 /mnt/data02 /mnt/data03 /mnt/data04

FAST_GET=0:

FAST_GET=0 ./start-buckit.sh /mnt/data01 /mnt/data02 /mnt/data03 /mnt/data04

Notes:

• MINIO_CI_CD=1 may be required when using root-disk paths for a control rig.
• Default credentials are buckitadmin/buckitadmin.

Warp Flow

Load once with fast path on:

./bench-warp-2m.sh load-on

Then restart the server with fast path off and run:

./bench-warp-2m.sh off

Then restart the server with fast path on again and run:

./bench-warp-2m.sh on

Cold TTFB Flow

1. Generate presigned URLs for the objects you want to probe.
2. Run mc admin trace --verbose from another shell.
3. From the host, run:

./cold-curl-ttfb.sh "<url-1>" "<url-2>" "<url-3>"

This prints one curl time_starttransfer value per GET.

Current Notes

The current single-HDD control results and interpretation are documented in:

• docs/single-trip-hdd-bench/README.md

That document records the June 4-5, 2026 single-disk control run, including:

• the 2 MiB positive signal
• the noisier 640 KiB runs
• the path dependency checks

The two-node real-HDD (d3.xlarge) cold-TTFB A/B runs are documented separately in:

• docs/single-trip-hdd-bench/two-node-d3-2mib.md — June 5, 2026 2 MiB run on a 2-node EC:3 cluster (300 pooled cold samples per arm: ~14% median TTFB reduction with the fast path), plus the method corrections that invalidate the earlier confounded 2 MiB numbers.
• docs/single-trip-hdd-bench/two-node-d3-640kib-ec2.md — June 5, 2026 640 KiB run with the set reconfigured to EC:2 (4 data + 2 parity). Crossover result: the fast path is worse at low/median percentiles (~4 ms floor from cancel-close overhead) but better in the tail (p90 −26%, mean −18%).
• docs/single-trip-hdd-bench/two-node-d3-profiling.md — June 5, 2026 source-level GET phase profiling (BUCKIT_FASTGET_PROFILE=1). Finds the bulk of a cold GET is cross-node cold disk opens (OFF pays the phase twice; ON's open phase is gated by the slowest-of-6 incl. discarded parity, plus a header→body barrier).