Re:Optane Memory – A bridge between DRAM and storage
by mrjones on Monday March 27, 2017 @02:32PM
(#13970400)
Actually, the linked paper discusses 112-115us of total read latency for NVMe SSD, not 14-15 (see Figure 5).
90us of raw device latency, then software overhead of 22-25us (2-3us driver, 6-7us kernel, 14-15us user space).
So, the 6us access latency quoted in the Anandtech article would be significantly faster than NVMe SSD.
Re:Optane Memory – A bridge between DRAM and storage
by convolvatron on Monday March 27, 2017 @02:18PM
(#13970263)
seems like it might be useful as a transaction log?
Re:Optane Memory – A bridge between DRAM and storage
by angry_octet on Monday March 27, 2017 @06:37PM
(#13972456)
The latency of a regular SATA SSD is more than good enough if you want a log of writes to spinning rust. The increased write endurance might be appealing for something like a ZFS intents log (ZIL).
Too small to provide much of an cache, when you can have 256GB of SSD for the same price.
It is a pity that the marketing docs don't address how this impacts on write consistency. Instant write sync is the primary advantage of hardware RAID, but still, how much better is this than a normal SSD?
Re:Optane Memory – A bridge between DRAM and storage
by Dylan16807 on Tuesday March 28, 2017 @01:30AM
(#13974340)
> The increased write endurance might be appealing for something like a ZFS intents log (ZIL).
These first generation products don't actually have better endurance, though.
You lose SSD transfer speed, while gain _some_ hardware response latency.
NVMe response latency (on hardware side, not OS) is around 14-15micro-seconds on reads, and tops out at 200micro-seconds on writes [1]. Which is nearly 2-3 orders of magnitude below spinning rust. While Optane can halve this to 7micro-seconds on reads. The price/pref vs established NVMe tech is non-existent.
When I can spend $X/GB to get a 2000% performance bump, its hard to justify spending $X*4/GB to get 2050% performance bump. Most applications/data centers have yet to transfer to NVMe. Offering marginal gains a sizable markup is a risky strategy.
Furthermore Optane is nearly 2 orders of magnitude lower than RAM in latency, and 5 orders of magnitude lower transfer. So the idea of "non-volatile RAM" has yet to arrive.
---
Also "Transparent to the OS" should strike fear in your heart. The OS's job is to manage hardware, memory, and caches. Why drop this?
[1] https://www.cs.utah.edu/~manua/pubs/systor15.pdf