No, I don’t mean to be bad, but the spinning HDDs’ prices will remain high even if the post-Thailand flood production has resumed to normalcy.
According to IHS iSuppli, a market research intelligence firm, the prices will continue to hold steady and will not fall to pre-flood level until 2014. The reason is simple. The prices of the hard disk drives are pretty much dictated by the only 2 real remaining hard disk companies in the world – Seagate and Western Digital. These guys controls more than 85% of the hard disk market and as demand of HDDs outstrips supply, the current hard disk prices are hitting the bottom line hard for just about everyone.
But the bad news is turning into good news for solid state storage devices. NAND-Flash based devices are driving a new clan of storage start-ups in the likes of Violin Memory, Kaminario, Pure Storage and Virident. The EMC acquisition of XtremIO was a strong endorsement that cements the cornerstone of all enterprise storage arrays to come. Even the Register predicted that the EMC VMAX will be the last primary storage array before the flash tsunami.
The NAND-Flash solid state of multi-level cells (MLCs) and single level cells (SLCs) and even triple level cells (TLCs) are going through birth, puberty, adolescent extremely fast because the demand for faster and faster IOPS, throughput and lower latency is hitting at full speed. And it is likely that all the xLCs (SLCs, MLCs and TLCs) could go through cycle in an extremely short lifespan, because there is a new class of solid state that is pushing the performance-price envelope closer and closer to speed of DRAM but with the price of Flash. This new type of solid state is Storage Class Memory (SCM).
First of all, we are all aware that Flash is a type of solid state storage, but DRAM-based storage can beat Flash in terms of speed, IOPS, latency and throughput. But unfortunately DRAM is memory and DRAM, just like memory, is volatile. Some sort of DC battery or super capacitor has to sustain the content of DRAM in order to keep the content stored in the DRAM before it is written to stable storage. DRAM is expensive.
SCM is a new solid state device that much lower latency and with speed closer to DRAM than Flash, but it is non-volatile. It will have low cost per bit and the aim is to get the price close to HDDs. It will also have a much higher reliability of millions of write cycles compares to tens of thousands in MLCs and TLCs.
There are many candidates for SCM. These include:
- FeRAM (Ferroelectric RAM) using Ferroelectric Field Effect Transistor (FeFET) technology
- MRAM (Magnetic RAM) with Racetrack memory technology
- RRAM (Resistive RAM) with organic and polymer memory technology
- Memristor (Memory Resistor)
- Solid Electrolyte
- Phase Change Memory (PCM)
Here is a glance of some of the emerging SCM memory technologies in the circa of 2008.
All these technologies were being created in 2008 and one of them, Phase Change Memory (PCM), is the most promising with IBM leading the way.
The promise of PCM is massive because it will be 100x faster than Flash, with access times in nanoseconds rather than microseconds in Flash, retention time of 2-10 years as opposed to milliseconds in DRAM (remember that DRAMs are volatile?), and will be cheap enough for mobile devices. Even better is PCM is word-addressable and very efficient. This means that if a read operations require 1 byte, it will read 1 byte. In conventional storage I/O access, a single block is the smallest granular unit of an I/O operations.
PCM in a nutshell is based on the ability to send controlled bursts of electricity to the memory device made out an alloy of Germanium-Antimony-Tellurium (GeSbTe) and nudging it to change its physical state or phases.
In June 2011, a team in University of San Diego already tested a prototype PCM module and it was able to run 7x faster than the most efficient Flash-based SSDs today. It was able to deliver 327 MB/sec of read throughput and 91 MB/sec for write throughput. Here’s a look at one of the prototype PCM modules below.
The solid state future looks bright thanks for the high prices of HDDs today. And eventually, the storage class memory devices will overtake Flash-based solid state devices of today.
Pingback: Expensive Hard Disk is goo « Storage Gaga
Great post! I share your enthusiasm about the implications of SSDs on storage tech, but 2 concerns come to mind:
1) The life cycle of these devices is a totally different kettle of fish. There are new failure modes, some of which are closely tied to usage patterns, which is related to my second point…
2) Higher up the stack, there is cause for consideration, vis-a-vis write amplification and the effects of the write cliff. This is practically a NoOp if you have well understood bounds on how your app will use the device, but that certainty goes out the window in a multitenant environment (in particular IaaS).
I guess the bottom line is I think it’s a provisioning headache.
Obviously solutions will emerge to these issues, but for now as far as I am aware these are open questions.
Thanks for your comments.
I agree with what you have said about reliability. There are concerns about the present NAND-Flash solid states, where there is a limited lifecycle in terms of write/erase. The failure patterns would be different and that is why, as you said, higher up in the stack, there are algorithms that spreads the writes to the devices. Shadow paging/copy-on-write filesystem is one algorithm that could alleviate the concern about distributing writes.
Write cliffs will happen, that’s the design of the NAND-Flash SSDs. Again, it is typically the file systems and its features of caching, garbage-collection or certain mechanism higher up in the stack to alleviate the effects of write cliffs. Overprovisioning is used by certain SSD controllers, and some controllers added a level 2 cache using DRAMs. Higher up, mechanisms to bunch up content locality could help. ZFS with Hybrid Storage Pools and its Adaptive Replacement Caching would allow associated blocks to be grouped for different Read and Write patterns.
In a larger context, it has become increasing clear that eventually the Cloud Storage in a multi-tenant environment will come to a point where the delivery of performance becomes part of the equation. Right now, capacity is more relevant, but unfortunately, most Cloud Storage solutions have to have the entire architecture revamped for SSDs. I have seen ZFS being pervasive in clouds like Joyent but I can’t comment about the significance of others.
Sadly, I believe that you are very right because storage vendors today are just throwing in SSDs to address present problems and not doing much to design something for the future and the things to come.
Thanks for your comments