[GTALUG] NUC NUC NUC
D. Hugh Redelmeier
hugh at mimosa.com
Mon May 20 11:42:00 EDT 2019
| From: Russell Reiter via talk <talk at gtalug.org>
| Optane was intended to be a cache memory to increase the performance of
| conventional spinning HD's under Windows OS. However, I've been booting
| Fedora from both a 500gb SSD and 32gb Optane Nvme as I tinker with my own
| desktop.
I think Optane is just a brand name and that brand name gets attached
to several different things, some of which are not yet being sold.
The original promise was: non-volatile memory that would get speed
close to RAM and price close to flash. (They also need better
durability than flash.) There's a big gap between those, and anything
in between ought to have a market. They just haven't been able to
accomplish much.
- they wanted to produce things that fit in RAM sockets. That meant
either
(1) no change from RAM interface (unlikely, but suggested-by-omission
in early marketing) or
(2) memory interfaces that were augmented to support the new protocols
(delivered with some new Xeons, I think)
- the NVMe stuff has had much worse durability than originally
promised.
- the NVMe stuff has had quite small capacity compared with most SSDs
- NVMe SSD has mostly been fast enough that the Optane stuff isn't
compellingly better
| Most certainly booting to a login prompt is fractionally quicker
| on the Nvme than on the conventional SSD.
That's what I'd expect. But I'll admit to no experience with Optane
and I haven't been following it closely.
| However recently I up-sized my
| Nvme and have populated my M.2 slots with a 250gb WD black Nvme for boot
| and now added an additional 1TB to the second M.2 slot with F29 still on
| the SSD. Copying a 100gb image to the 1TB drive really hit performance tho
SSDs come with different performance trade-offs. Most inexpensive
SSDs have (on-board) controllers with only small amounts of RAM. This
makes them slow down a lot after a modest burst of intensive writing.
That's a fine trade-off for many of us but not for all workloads.
You can pay more and get SSDs with enough RAM to not have this
performance problem. I don't know enough to give specific advice.
I have recently learned that some cheap NVMe drives can ask the OS to
allocate system RAM for the exclusive use of the controller. This
isn't a conventional data cache, but something much stranger. It's
called "Host Memory Buffer" (HMB). I think that vendors don't explain
it because they think consumers won't understand it.
HMB might be a great trade-off, or a horrible hack. I don't know.
What happens when the power fails? Or when the system crashes /
reboots?
Recent Linux and Windows support HMB. I assume that UEFI firmware does
not. So use of HMB must be an optional speed-up.
| and that was probably due to the lack of a decent heat sink.
| I just ordered a hteatsink fro the internal 1TB
Why do you think that this was heat-related? It might be, but that
would not be my first guess. (I am not an expert on this.)
| Perhaps with the NUC form factor heat might
| be a problem on a larger sized Nvme but with USB-C you have wiggle room for
| adaptation.
The NUC form factor certainly reduces the heat disposal and thus
limits components. But the main such component is the CPU. I've not
heard of it being a problem for consumer 3.5" SSDs (what Evan intends
to use) or NVMe drives.
More information about the talk
mailing list