IO_uring Squeezes More Performance With Linux 5.13

[phoronix]

Phoronix: IO_uring Squeezes More Performance With Linux 5.13

Merged as part of the block subsystem changes for the Linux 5.13 were the usual assortment of enhancements to the exciting IO_uring. With this next kernel there is yet even better performance out of this morning Linux I/O interface...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.13-More-IO_uring

[arthureroberer]

based

[onlyLinuxLuvUBack]

So there are more performant IO_uring squeezes in 5.13 but for which drive type?
Will only nvme drive see this ?

[coder]

So there are more performant IO_uring squeezes in 5.13 but for which drive type?
Will only nvme drive see this ?

The increase in IOPS will surely be limited to NVMe drives, as SATA devices have much greater transaction overhead and maxed out in the ballpark of 100k IOPS.

But 3M IOPS is something not even most NVMe drives can manage. I didn't do much digging, but the pinned tweet on his feed that highlights his 5.9 performance numbers (which are similar and slightly lower, suggesting probably the same hardware) specify the storage device as being: "Single Gen2 Optane SSD".

Here's the post linked by that tweet: https://lore.kernel.org/io-uring/4af...kernel.dk/T/#u

Presumably, the rest of us will simply experience this change as a very small reduction in CPU utilization, for those apps that use io_uring in the targeted fashion.

[pal666]

So there are more performant IO_uring squeezes in 5.13 but for which drive type?
Will only nvme drive see this ?

do you know any nvme drive capable of 3m iops? i'm pretty sure peak performance is for cached case

[S.Pam]

When the disk is the bottle neck then these improvements mean less CPU overhead, and that is also good

[coder]

do you know any nvme drive capable of 3m iops? i'm pretty sure peak performance is for cached case

In the link I posted above, he specified that he's using a "Single Gen2 Optane SSD". He further stipulates:

"Workload: Single thread random 512b O_DIRECT reads"

O_DIRECT means uncached, for those who didn't know.

Intel claims this Optane SSD can achieve a whopping 4.6M IOPS of 512-byte random reads:

https://www.servethehome.com/new-int...tes-pcie-gen4/

[pal666]

https://www.servethehome.com/new-int...tes-pcie-gen4/

i was waiting for upgrade for 905p, but it looks like i'll have to go with flash

[coder]

i was waiting for upgrade for 905p, but it looks like i'll have to go with flash

Yeah, I almost pulled the trigger on their now-discontinued 800P. But, it was only PCIe 3.0 x2 and still rather pricey for its capacity.

I am hoping they offer something comparable, in PCIe 4.0 and their next-gen Optane, but I think it's looking unlikely.

One upside to this stupid Chia crypto-currency is that when it crashes, there should be a lot of big, relatively inexpensive SLC NVMe drives available.