Moore’s Law (and family) are a significant boon to anyone writing distributed systems. As the systems grow over time, so too does the underlying hardware get more powerful, allowing exponential growth to be readily managed. Programs need to be modified to cope, of course, but the raw compute power is there and available. That said, designing algorithms that can scale with more than an order of magnitude of growth remains a difficult challenge, and regular rewrites are inevitable. In the interest of minimizing that number, here are a few trends I’ve observed that should continue into the next few years. Caveat emptor, this is my personal perspective, but if it doesn't line up with your experience I'd love to hear about it.
1. Compute Units are Static
- Moore’s Law tells us that the number of transistors on a die double every 2 years, and practical experience tells us that performance doubles every 18 months. In the multicore era, however, the performance of a single core has largely plateaued, with further advances coming in the form of more cores per chip.
- RAM and on-die caches roughly keep pace with CPUs, which means that now that we’re into core count growth, the RAM and cache per core is now approximately fixed.
- Disk storage is also keeping the same pace, and hence disk per core is now also approximately fixed.
Thus, we can define a theoretical “compute unit” as, say, 1 core, 2GB RAM, 20GB flash*, 200 GB disk (adjust numbers as you see fit); this unit should stay roughly stable going forward, with the number of such units increasing exponentially.
Note that the purchasable range of resources-per-core is generally increasing; this represents the most economical average rather than a strict upper or lower bound. Resources do not need to be apportioned to each process in these ratios, either; e.g. disk-heavy and CPU-heavy processes can coexist on a machine and balance each other out. But it's worth looking closely at anything expected to grow significantly on one axis and not in others, to see if it's sustainable long term.
* Flash adoption remains spotty, but early indications put it roughly halfway between RAM and disk on performance and price, and it seems to be keeping up the growth pace. I speculate it will become a standard inclusion going forward, but this isn’t as clear as the others.
Update: One of my colleagues has pointed out that disk performance is not keeping pace with capacity increases (neither seek nor transfer rates). I'm not sure how to include this detail in the compute unit definition itself, but it will likely prove to be significant and help drive the adoption of flash going forward.
2. The Datacenter is the New Machine
This is already largely true, but the trend will only get stronger. This is driven by two main trends:
- NUMA architectures are reducing sharing and performance uniformity within machines, and
- Intra-datacenter connectivity is growing faster than individual machine performance, reducing the impact of communicating between machines and racks.
Thus, the trend of running programs as many smaller components that may or may not share any particular physical hardware will continue to get stronger. Correspondingly, virtualization will be increasingly necessary for isolation, further abstracting processes from the hardware they run on.
We may also start to see this trend spread to the client-side as well, but I’m not sure what form it would take. In-home compute sharing?
3. Decomposable Programs are Ideal
As a corollary to the previous point, algorithms that can be decomposed into fixed-sized chunks are best. They can be apportioned among CPUs/machines/racks easily, and scaling is decoupled from machine growth (which occurs in generational bursts when hardware is refreshed). Which means that if sharding is necessary, dynamic sharding - a variable number of fixed-size chunks - should be preferred over having a fixed number of shards that grow separately.
MapReduce (or Hadoop), Pregel (or Giraph), and GFS (or HDFS) are all examples of decomposable programs, tackling data processing, graph computation, and storage, respectively.
4. Consumer Bandwidth is Lagging
Consumer network growth is lagging behind intra- and inter-datacenter networks, and hardware trends. Some of this is due to a shift towards mobile, where equivalent resources (primarily images and videos) are smaller, but it holds on wired internet as well. It is unclear to me how long this will last - gigabit customer internet is already being tested in a few places, for example - but for the time being it remains true. This means inter-datacenter communication is growing cheaper relative to communicating with consumers. While it remains true we should expect more work to be pushed remotely to the “cloud” and an increase in datacenter entities cross communicating, and a slow shift towards smaller and more expensive resources (heavier compression, more computationally expensive codecs, etc) for customers.
5. Latency Doesn’t Scale
A fact that has always remained true: the speed of light is (effectively) constant, and does not decrease in line with other improvements. As bandwidth increases latency will continue to dominate long-distance performance, and round-trip times will continue to constrain serial throughput. Together, these will continue to push a couple of long-lived trends:
- Reducing data proximity. Caching (local and edge) and having data prefetched are key in reducing latency, and get comparatively cheaper over time. One could imagine future efforts that will proactively push resources straight into user networks or devices as well.
- Supporting long-lived connections and message parallelism. Connection establishment requires at least one round trip, and any form of serial channel use cannot fully utilize one. Together these will lead to more connection sharing where possible, and more long-lived idle connections otherwise.
IEEE Industry Connections Ethernet Bandwidth Assessment
(Thorough study of bandwidth trends at all levels)