32 bit architectures and storage speeds

Russell Doty rdoty at redhat.com
Fri May 3 14:45:06 UTC 2013


On Fri, 2013-05-03 at 10:54 +0200, Jan Safranek wrote:
> Hi,
> 
> I'm implementing Block Server Performance Subprofile and one of the
> values it reports is nr. of sectors read/written from/to a block device.
> 
> On 32-bit architectures, these values are provided by kernel with 32-bit
> precision. If someone is using iSCSI over 10Gb/s link, the 32 bit
> counter overflows in roughly 1600 seconds. Therefore my provider should
> periodically detect this overflow and recalculate 64 bit number out of
> it. It's not so difficult, but:
Good question. RHEL 7 is planned to be 64 bit only, so this shouldn't be
an issue there.

What happens when you overflow? Does the counter wrap, or do bad things
happen?
> 
> - Do we need to support 32 bit architectures at all?
I would say minimal support. It would be nice if it worked, but I
wouldn't worry too much about a counter wrapping. If the results are
reliable over a 10 sec to 60 sec window, I wouldn't worry about it too
much.
> - What is the maximum read speed we need to support? I need to set the
> timers right, so I don't miss an overflow of the counters, and still not
> take too much CPU. Is it 10Gb/s? What will be iSCSI/FC/SSD speed in 10
> years? Remember, the 32-bit machine might be virtual, running on really
> powerful HW.
In 10 years we will be approaching Tb/s speeds. Fibre channel and
Infiniband are at 40Gig today, and 100Gb/S Ethernet is emerging. Tb/s
fabrics are likely in this timeframe. 
> 
> Jan
> _______________________________________________
> openlmi-devel mailing list
> openlmi-devel at lists.fedorahosted.org
> https://lists.fedorahosted.org/mailman/listinfo/openlmi-devel




More information about the openlmi-devel mailing list