Glibc's Per-Thread Cache Is Helping Out Some Benchmarks

[phoronix]

Phoronix: Glibc's Per-Thread Cache Is Helping Out Some Benchmarks

Released this week was the shiny new glibc 2.26 GNU C Library with the notable new feature being the per-thread cache for malloc...

http://www.phoronix.com/scan.php?pag....26-Redis-Test

[bug77]

Nice. Specific solutions/optimizations are not expected to improve things across the board. As long as they don't introduce regressions, they're golden.

[ayumu]

It wouldn't hurt to include musl in these.

[iptriz]

musl is most likely going to be an absolute performance disaster. But that can be motivational.

[arjan_intel]

It wouldn't hurt to include musl in these.

just tested this... had to disable key security features in the compiler flags (FORTIFY_SOURCE and stack protector) but even with that, a musl-enabled redis runs at roughly half the performance of the glibc-2.26 redis (range is from 1.8 to 2.2 or so)

[Mark Rose]

Nice. Specific solutions/optimizations are not expected to improve things across the board. As long as they don't introduce regressions, they're golden.

Sometimes it does. The benefit here is the ability to alloc without a lock. The disadvantage is increased memory consumption in heavily threaded applications, such as Java, especially on 32-bit systems where address space is constrained.

[arjan_intel]

Sometimes it does. The benefit here is the ability to alloc without a lock. The disadvantage is increased memory consumption in heavily threaded applications, such as Java, especially on 32-bit systems where address space is constrained.

at least the heavily threaded applications is also directly where the benefits pay off (so it's not that you pay a price without benefits, they're at least correlated)

do 32 bit systems still exist (where java and the like matter) ? Even android went 64 bit

[bug77]

Sometimes it does. The benefit here is the ability to alloc without a lock. The disadvantage is increased memory consumption in heavily threaded applications, such as Java, especially on 32-bit systems where address space is constrained.

Didn't say it doesn't happen, just that it's not a sensible default expectation from these kind of things.

[caligula]

Sometimes it does. The benefit here is the ability to alloc without a lock. The disadvantage is increased memory consumption in heavily threaded applications, such as Java, especially on 32-bit systems where address space is constrained.

You must be joking :O Heavily threaded applications and 32-bit. Wut. In 2017. Yes, it's 2017.

[caligula]

Maybe someone with more practical wisdom could explain what these tests do and how do they relate to real world web applications.