On 8 Oct 2026, at 04:06 UTC, Michael Paquier wrote: > I was wondering about some micro-benchmarking to > reduce the amount of noise to zero
Yes, and for the replay results there is one specific thing to isolate: with DYNAMIC_BMI2, ZSTD_initDCtx_internal() does CPU feature detection via CPUID for every new context, in both 1.5.5 and 1.5.7 [0]. Context reuse avoids that as well as palloc/pfree. My benchmark host is a KVM guest, so I wonder whether the cost of that check contributes to the difference. This is still only a hypothesis. My benchmark hosts are unavailable today due to data center maintenance. When they are back, I can measure context creation and library versions. Best regards, Andrey Borodin. [0] https://github.com/facebook/zstd/blob/v1.5.7/lib/decompress/zstd_decompress.c#L252-L275
