This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Add Prefer_MAP_32BIT_EXEC for Silvermont


On 12/11/2015 10:25 PM, H.J. Lu wrote:

> 3% speedup is for my typical workloads, which is running GCC.   For
> an artificial benchmark, I got
> 
> Old glibc:
> 
> [hjl@gnu-slm-1 dlcall]$
> /export/build/gnu/glibc/build-x86_64-linux.old/elf/ld-linux-x86-64.so.2
> --library-path /export/build/gnu/glibc/build-x86_64-linux
> intel64/dlcall
> 
> Time for 1000000 calls into dynamic library   1.00   ,  22.83 MT,
> 22.83 MT,  22.83 MT,       0 T
> 
> New glibc:
> 
> [hjl@gnu-slm-1 dlcall]$
> /export/build/gnu/glibc/build-x86_64-linux/elf/ld-linux-x86-64.so.2
> --library-path /export/build/gnu/glibc/build-x86_64-linux
> intel64/dlcall
> 
> Time for 1000000 calls into dynamic library   1.00   ,   2.87 MT,
> 2.87 MT,   2.87 MT,       0 T
> 
> which is 8X speedup.

Ouch.  Can't you label those chips as 32-bit-only?

What does your benchmark look like?

Can you run it with PIE?  And please also measure the performance of
gettimeofday and sched_getcpu.

I suspect you need kernel support to avoid the performance hit.

Florian


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]