VMware imagines 'memory servers' – a new source of shared software-defined RAM VMware is working on a software-defined memory effort and thinks it could lead to the creation of "memory servers" – boxes full of memory that can be shared across a cluster. DRAM can account for 40 per cent or more of a server's cost, but remains expensive. Memory is also in demand, as more workloads require in-memory …
With non volatile storage having quicker access and transfer rates than RAM of a few generations ago I’d need to see the numbers to understand the benefit of accessing off system RAM across some interconnection vs local sad. If it’s quick enough to not be noticeable then it’s a great idea.
An extension could be to pool all the ram under management (imagine 12 esxi hosts each with 200 GB ram with some hosts needing more than others) so existing hardware can be used to prove the concept with an eye for dedicated vsphere ram server later.
Would be a winner for bladed systems.
One one side of reality we carefully organize VMs to not cross numa node boundaries because the interconnect between 2 CPUs slow down memory access and performance too much... And on the other side, these lunatics think "off-system in the next rack" would be useful for any practical application.
Just from glancing at product brochures, I would guess that it's on the order of at least 2-10 times worse memory latency (looking at latency claims in the brochures for RDMA over infiniband vs latency claims for locally attached DRAM, which seem to be about 1us-ish vs about 200ns-ish) and the bandwidth is probably much, much lower. The usual effects that make small CPU caches useful probably come into play a bunch (programs often spend a lot of time accessing a working set much smaller than the whole of main memory).
VMWare aren't the first people to suggest that memory disaggregation might be a good idea sometimes. Anyone who's benefitted from using memcached in clustered mode (which, I guess if you squint real hard, can be looked at a bit like a slow application level implementation of memory disaggregation) in production definitely knows it. :)
Vmware has been punting its terrible readwrite IO problems on its latest vsphere7.x series that seem to enjoy excessively killing lowend sdcards and ssds by simply claiming not to support them anymore. I wonder if the issue is going to come back to bite vmware in the rearend sooner than later.
I wonder if the issue is going to come back to bite vmware in the rearend sooner than later.
Nah, they've been going long enough to get vendor lock-in on their side. As long as the annual cost is less than what it would cost to switch to a different hypervisor, they've got clients for life.
See also, Oracle, Microsoft, etc.
TCO savings of 30-50% with a performance hit of ~15%? In my mind, maybe not so much. It is important to remember that HW contribution to TCO (HW, SW, Service, Personnel, Power, Space...) is relatively small. Saving even 50% on something that is, maybe, 15% of TCO may not be worth a 15% hit on performance. Especially since the implementation of 'memory servers' won't be free. Huh?
"If that gives you pause, VMware said its tests indicate total cost of ownership savings for hardware of between 30 and 50 per cent can be achieved. "
I can't wait to use this to extend the memory in my tape server. Perhaps then, the tape drive will be able to compress the files so it can fit 400% more data on a tape instead of the meager 200% it claims on the box. Because vendor claims and numbers never lie, especially when a percent sign is involved...
To be fair to VMWare, virtualised memory as in swap. paging et al is not what they are claim IMHO. What they are talking about is virtualising physical RAM capacity. In short, doing what IBMs hypervisor has done for years, managing AIX (and now Dead Rat probably) instances hosted on Power hardware where the physical RAM can be pooled into groups such that the virtual servers can "borrow" unused RAM from other servers in the same pool. Maximum and minimum RAM configurations are set to limit hogging. In short, X86/x64 getting what real OS have had for years. In commercial operations, slowing the developers so the bean counter end of season/financial year runs can go faster is useful.. Something the old IBM did well.
They're talking about memory over network and Optane, so it's swap.
It sounds like the real product is:
- Fancy swap limited to containers so your software-defined-swap components don't need to be swap-safe themselves. That's freedom from a hefty limitation.
- A tool to gather up all your random bits of unused fast storage and pool it for container swap.
It seems like this product is only practical on systems with incredibly fast interconnects. Even so, the chance for complex patterns to appear in many simple components interacting with each other would keep me up at night. I'd hope for The Reg to test it for us but this makes me think they don't have that kind of equipment in-house.
Depends on OS, hardware and language in use I suppose. COBOL move SPACES to DATAROW sort of thing or rely on hardware or OS to clear memory contents before reallocation. (unlikely IMHO, anyone care to clarify this?)
As for swap over network, so VMWare are doing same as Suns diskless workstations did , oh about early 1990s ? Progress ? Some concepts keep being re-implemented.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
