New Linux Kernel Patches From Intel Delivering +18% Database Performance
In addition to the recent Linux kernel patches out of Intel for Cache Aware Scheduling for better performance, separately, another interesting new patch series was sent out this week for the Linux kernel. The patches rework some low-level Linux kernel memory management code and at least for database workloads the early benchmarks are showing possible 14~18% faster database performance with PostgreSQL.
Intel Fellow Thomas Gleixner of Intel-owned Linutronix posted an interesting patch series this week. The 19 patches are a rewrite of Linux's memory management subsystem's memory-mapped concurrency ID (MM CID) code.
Long story short for those not concerned about the low-level, inner-workings of the Linux memory management code but the end-user impact, the early benchmarks of this are showing a very positive improvement:
Up to 14% faster PostgreSQL performance with sysbench and an additional 3% gain also possible with another code branch being tested.
In a thread-create teardown micro-benchmark was as much as a 30% improvement over the upstream Linux kernel code.
Additional performance benchmark data from these code branches are being sought for seeing other possible gains or areas for improvement still. If I can find the time, I'll be spinning up some benchmarks of the proposed code soon.
Those interested can find the work via this Linux kernel mailing list thread.
Intel Fellow Thomas Gleixner of Intel-owned Linutronix posted an interesting patch series this week. The 19 patches are a rewrite of Linux's memory management subsystem's memory-mapped concurrency ID (MM CID) code.
Long story short for those not concerned about the low-level, inner-workings of the Linux memory management code but the end-user impact, the early benchmarks of this are showing a very positive improvement:
Up to 14% faster PostgreSQL performance with sysbench and an additional 3% gain also possible with another code branch being tested.
In a thread-create teardown micro-benchmark was as much as a 30% improvement over the upstream Linux kernel code.
Additional performance benchmark data from these code branches are being sought for seeing other possible gains or areas for improvement still. If I can find the time, I'll be spinning up some benchmarks of the proposed code soon.
Those interested can find the work via this Linux kernel mailing list thread.