I'm looking at an x86 testcase with a 16byte overaligned struct and new isn't using github.com/facebook/zst

argh wrong paste

operator new(unsigned int, std::align_val_t)

is there a g++ option to make is use aligned new?

000c3460 T operator new(unsigned int, std::align_val_t)@@CXXABI_1.3.11

-std=c++17 -faligned-new don't change anything

which gcc are you using?

gcc (OpenIndiana 13.3.0-oi-0) 13.3.0

you can try 14, it is also available on oi

same problem

no idea then. I havent tracked gcc that much and I guess my google is just as good as yours:D

Well, that's a fail. My test box hangs drive IO under load. Ye olde LSI 1068E in IT mode, it just dies, no kernel messages or anything.

Kurlon_ which and how many disks?

sata?

8 x Samsung 850 Pro 128GB SSDs

SATA

I'm seeing lots of reports of that being an issue on these specific cards.

I see...

So, plan B is back to an IBM M5014 that'll get flashed to IT mode for 4 of the drives, 4 more off the onboard intel chipset SATA with an unhappy bend in the connector.

Or... not, no IT mode firmware for this antique.

For $30 on eBay plus some updated SAS cables I can get a 9300-8i, feels like the way to roll.

For now, M5014 setup with four raid 0 virt drives, one for each SSD, other four on the chipset SATA, seems to be behaving OK.

Actually getting a bit better throughput with things split over two controllers.

I have no idea about the GCC thing above either, but C++ choices are usually mystery to me

w

Interesting side effect of this bad storage setup, during a scrub the 'virt' drives report an IO count 10 times what the bare SATA drives do, but the total k/s matches.

Kurlon_: 4k vs 512 byte sector sizes, perhaps?