DISQUS

GoatGuy · 11 months ago

On the one hand, solid-state drives are a revolution ... on the other, they're not addressing one of the key bottlenecks that is today impacting the ability of modern processors to process ''more, faster, in parallel''.

The revolution is easy enough to see: a device modelled after the conventional hard drive, supporting the protocols of the SATA interface standard allowing it to be dropped in to most any computer design sporting the SATA channel. ... Gigabyte sizing that is similar to the capacity of conventional hard drives. ... Superior 'data access' performance - at times vastly superior - and near saturation of the SATA 300 megabyte per second capacity. Anecdotally, for otherwise identical laptops, one with a top-of-the-game conventional HD, and the other with a top flight SSD ... the SSD version boots in 1/4th the time, programs load between 1/10th to 1/2 the time, and to the human tester, the SSD version seems hardly ever to 'bog down', as is so common for conventional HD based units.

The shortcoming though (and really this is no fault of any of the SSD makers, but more of the industry as a whole) is that the channel - SATA II - is itself pathetically underpowered compared with modern DRAM 'performance * capacity'.

This is what I mean ... on a little stroll through computing history.

Early personal computers were uncontravertibly DRAM challenged. DRAM moved data in byte or 2-byte words, was clocked at 8 MHz and attained data rates of about (4 x 2) = 8 megabytes per second. The total DRAM was also small, from tenths to a few megabytes. Hard drives at the time could deliver 2 MB to 5 MByte per second ... so theoretically all (1 MB) of DRAM could be swapped onto and off of a HD in the high milliseconds to a couple seconds. This in turn allowed many programs to be split up into 'overlays' and other schemes that could swap modules out in time spans short enough that the attendant user saw only tiny delays.

To critical people (humbly, like me), it looked essentially smoothly virtual.

However, due at its core to Moore's Law (of transistor doubling every 18 months), inevitably the amount of DRAM (especially) installed in computers grew at an exponential rate. Today for instance, it is not uncommon to buy laptops and desktop machines running Vista with 2,000 MB to 4,000 MB of memory. Even more so, the Apple Mac community regularly purchases desktop machines with 16,000 MB to 48,000 MB of memory. Video, heavy graphics ... uses all of it too!

Continuing the original analysis ... and assuming that we all in the future will have 16+ GB of memory on our machines, the question is, how fast do current top-of-the-line nonvolatile memory devices (HDs and SSDs) perform?

Conventional HDs are endlessly benchmarked, and are shown to deliver 50 MB/sec to just over 100 megabytes per second in the streamed mode. Lets say "60 MB/sec" for the average HD. Well the 3,000 megabytes of a new laptop could swap in ... 2 x (3000/60) = 100 seconds! A couple of mintues! I trust the point is obvious: that the laying down of serial bits - even at blistering gigabit per second rates - still is pathetically slow, compared to DRAM capacity. The ideal of having a modest amount of memory, and a REALLY much faster offboard system to virtualize it to a larger pool ... isn't achieved, because the 60 MB/sec data rate is just way, way too small to make the transfer of thousands of megabytes possible "in a flash".

The SSD's appear to get closer to the ideal - which is pretty obvious: saturating the design limits of the transfer channel (in this case SATA). The above announced SSDs are able to deliver 250+ megabytes per second - a 4x improvement over the 60 MB/sec of a conventional drive - and have tons of buffer RAM to boot - to mitigate the relatively slower random WRITE performance. A great end. But it is really SATA II that is the limit.

What is needed is an interface that can mimic (or attain) the performance of the DRAM-to-CPU channel itself, these days on the order of 12 GB to 24 GB per second, interleaved. The performance of the DRAM-to-SSD channel should be at least 25% of the DRAM itself, along the lines of the 'DMA' access method sported in now-ancient computer designs. At least that way, one could swap meaningfully large fractions of a computer's entire DRAM to and from the drive, bringing back the kind of eyeblink virtual memory performance that made virtual memory so useful in days of old.

And that would mean performance of the channel in excess of 10,000 MB/sec (10 GB/s), or 30 times today's 300 MB/sec performance.

For conventional hard drives (which to me haven't yet given up the ghost), having read-write heads that access 32 to 64 adjacent tracks in parallel, simultaneously, would kick up the streaming performance to match the "new channel" capacity of 10,000 MB/sec. It would be possible to swap sizeable (25%) fractions of one's 4 GB memory to and from even conventional hard drives in the 'mid millisecond range'. Tenths of a second. "Eyeblink" rate.

Insofar as the hands-on feel would go, such performance, coupled with an economical amount of memory (say 10% of the machine-cost) would give pretty pleasing and ostensibly spectacular performance. I think that chasing the rainbow of 'ever more DRAM' should slow down, but the emphasis instead be placed on radically upgrading the nonvolatile memory (HD, SSD) channel.

POSTSCRIPT

I'm sure our technologist readers will counter that the solution really is to get rid of virtual memory and just have a large enough block of real DRAM to avoid swapping altogether. Yes, this is so ... but it is also more tightly limited to the physical design of the computer than the comparatively humungous capacity of nonvolatile drives.

Virtual memory's chief benefit (when fast enough) is that it doesn't impinge (or shouldn't) on application RAM performance when RAM demand is well below physically "free memory", and that it impinges "reasonably" on high-memory use applications (or many running simultaneously) when the aggregate demand "reasonably" exceeds physical memory. My point was to showcase that the present day streaming performance of Virtual Memory has lagged egregiously behind both DRAM bandwidth and gigabyte capacity, rendering it only useful for the smallest excursions of virtual performance.

Enhancing the performance of virtual memory to where the VM "channel" is on the order of DRAM's performance itself would restore VM as a key viable technology upon which much larger memory footprints could be sustained essentially with only small degradation in application performance and responsiveness.

GoatGuy