Build CUDA benchmarks once, but run in parallel#8489
Conversation
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
Merging this PR will degrade performance by 14.25%
|
| Mode | Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|---|
| ❌ | Simulation | take_10k_random |
197.9 µs | 255.8 µs | -22.63% |
| ❌ | Simulation | take_10k_contiguous |
218.5 µs | 276.4 µs | -20.94% |
| ❌ | Simulation | patched_take_10k_contiguous_patches |
232.2 µs | 291 µs | -20.18% |
| ❌ | Simulation | patched_take_10k_random |
244.2 µs | 303 µs | -19.41% |
| ⚡ | WallTime | cuda/bitpacked_u8/unpack/3bw[100M] |
352.6 µs | 299.3 µs | +17.8% |
Tip
Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.
Comparing adamg/unifrom-codspeed-gpu-build (9ed22a8) with develop (d020924)
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
AdamGS
force-pushed
the
adamg/unifrom-codspeed-gpu-build
branch
from
d9b785e to
d400751
Compare
0ax1
left a comment
0ax1
left a comment
There was a problem hiding this comment.
Looks good! Might be worthwhile to comment in the yml why we do this.
2 participants
Summary
Reducing the amount of GPU machines we need to get from AWS, and save some money.
An easy follow up here is to build on a machine without a GPU, which should also save some time on startup.