5 ms·
This was a lot of fun, one of the things that doesn't get much air time is that back in the early 2000's when "clusters" and "NUMA SMP machines" were competing
by ChuckMcM 9y ago
This was a lot of fun, one of the things that doesn't get much air time is that back in the early 2000's when "clusters" and "NUMA SMP machines" were competing with each other the big argument for large SMP iron was ACID compliant SQL databases like Oracle. Now that Google has implemented an ACID compliant SQL database across clusters it puts the final nail in the argument (for me at least) that "Some things only work on SMP machines"
- Pyxl101 9y agoDistributed ACID transactions across multiple DBs or clusters has been possible for a while if you have been willing to pay the latency penalty for the transaction to complete, hasn't it? As far as I know, Google systems have to pay this latency cost too. From an analysis of Spanner and Calvin [1]: > The cost of two-phase commit is particularly high in Spanner because the protocol involves three forced writes to a log that cannot be overlapped with each other. In Spanner, every write to a log involves a cross-region Paxos agreement, so the latency of two-phase commit in Spanner is at least equal to three times the latency of cross-region Paxos. I suspect that Google's innovation in making these cross-region or cross-cluster transactions viable is partly in their network: having an extremely reliable network with consistent low latency, a lot of which allegedly uses first-party network fiber between disparate locations. This is touched on in a paper that Google published [4]: > Many assume that Spanner somehow gets around CAP via its use of TrueTime, which is a service that enables the use of globally synchronized clocks. Although remarkable, TrueTime does not significantly help achieve CA; its actual value is covered below. To the extent there is anything special, it is really Google’s wide-area network, plus many years of operational improvements, that greatly limit partitions in practice, and thus enable high availability. I'm also not sure that technologies like Spanner necessarily address the same use-case as SMP machines. Aren't those use-cases concerned with achieving the highest possible transaction throughput for transactions that are potentially highly contentious and require low latency? Systems that solve that problem today still largely look like Oracle DBs from 2000 -- or a fleet of replicas reading executing the same state machine on a transaction log. Transactions against Spanner likely have high latency compared to traditional DBs[2]: > On the downside to Spanner, Zweben noted, that the price of Spanner’s high availability is latency at levels unacceptable in some applications. If you’re executing banking transactions, like in airline reservations, and you can afford a minute of delay while you make sure they’re committed in multiple data centers, Spanner’s probably a good database for that, he said. > Zweben says, however, if you’ve got a customer-service app that requires you to be extraordinarily responsive over millions of transactions and at the same time analyze profitability of a customer and maybe train a machine learning model, those simultaneous transactional and analytical requirements are better for a database [...] One benchmark I found suggests that read and update latencies can be quite high [3]. The graphs are missing units, but it seems to suggest update latency is higher than with traditional DBs. I've had difficulty finding other benchmarks just now, but the one other I found reported the same finding [5]: > An interesting pattern above is that queries 14-18, which are all updates, perform with higher latency on Spanner than the easy selects and non-bulk inserts. It's definitely innovative technology in that it allows you to scale read-oriented SQL workloads to a far greater degree, with better availability, but from what I've read it does not tackle all of the same problems and use-cases that traditional relational DBs are best at solving: high transaction throughput and low latency for contentious data sets. [1] https://fauna.com/blog/spanner-vs-calvin-distributed-consistency-at-scale https://fauna.com/blog/spanner-vs-calvin-distributed-consist... [2] https://thenewstack.io/google-cloud-spanner-view-field/ https://thenewstack.io/google-cloud-spanner-view-field/ [3] https://www.nuodb.com/techblog/benchmarking-google-cloud-spanner-cockroachdb-nuodb https://www.nuodb.com/techblog/benchmarking-google-cloud-spa... [4] https://research.google.com/pubs/pub45855.html https://research.google.com/pubs/pub45855.html [5] https://quizlet.com/blog/quizlet-cloud-spanner https://quizlet.com/blog/quizlet-cloud-spanner
- peterwwillis 9y ago> I suspect that Google's innovation in making these cross-region or cross-cluster transactions viable is partly in their network I think this understates it. Production distributed systems suck balls with unreliable infrastructure, and anyone who has ever tried to do realtime replication of all of their data (and lots of it) across the internet knows how completely unrealistic it is. It's like running a power cord across a highway to power a refrigerator.
- raarts 9y ago> It's like running a power cord across a highway to power a refrigerator. ??? Color me baffled.
- kuschku 9y agoIndeed, the trend is clear, but for now, nothing has changed. Google's implementation is not very helpful to all those companies, individuals, NGOs, and governments that have to follow privacy laws, HIPAA, etc though, because Google's implementation isn't open source, and these entities can't use Google Cloud. Or don't want to. Until we get an open source solution for this, SMP machines will be useful. And even then, you save money by having less and larger machines in your cluster than just having tiny ones. Larger machines means your overhead is reduced.
- jabl 9y agoThere's cockroachdb which is roughly open source spanner you can deploy on your own systems.
- kuschku 9y agoWhich has still horrible join performance, and several other tradeoffs. (see https://www.cockroachlabs.com/blog/cockroachdbs-first-join/ https://www.cockroachlabs.com/blog/cockroachdbs-first-join/) Maybe in a year, or two. But not today.
- phaemon 9y agoThat blog post is from a year ago. And since Cockroach DB 1.0 was only released in May this year, it's a bit misleading to link to that post as though it was the current state of the sofware.
- kuschku 9y agoThat blog post is referenced in their FAQ today, under the topic of what it can't do right now. Sorry if I misunderstood the situation, I’d appreciate any updated links. See: https://www.cockroachlabs.com/docs/stable/frequently-asked-questions.html#when-is-cockroachdb-not-a-good-choice https://www.cockroachlabs.com/docs/stable/frequently-asked-q...
- 9y ago
- 0xbear 9y agoI actually thing "NUMA SMP machines" are about due for a renaissance with the advent of extreme multicore CPUs. Next year we will see CPUs that run 32 cores (64 hyperthreads) per socket, and don't cost an arm and a leg. Current max is 28 cores per socket, at a staggering $13K per chip. With relatively economical 64 core / 128 thread options with nearly unlimited RAM capacity appearing, a lot more workloads will "fit". Needless to say, single-node systems are much easier/cheaper/faster to get right.
- ChuckMcM 9y agoThe irony of clusters of many core NUMA SMP machines is not lost on me :-) If things follow the previous patterns that will open up the market to a single/dual core, large memory, machine with a high speed networking and storage ports that is lower cost and lower power.
- peterwwillis 9y agoSome things do only work on SMP, because for some workloads no network fabric (that I know of) is as fast as the system bus. It's the classic Beowulf vs Mainframe problem: some things can be split up into discrete parts, and some things have to share so much that bandwidth is the bottleneck. Then you go from SMP to NUMA because your bus is too slow. But both are still faster than most (all?) networks. A lot of this also centers around how apps are designed. If you can rewrite it you can pick your architecture. But it's hard to take a pre-existing one and slap it on a new architecture and make it just scale. Mosix attempted to have a magical SSI that just scaled single pre-existing apps (even threaded shared-memory ones) across machines with no extra work, but it's got such a hard set of problems it never caught on. Now we just add intermediate abstractions until we can jam an app into whatever architecture we have.
- ChuckMcM 9y agoIt is exactly the Beowulf vs Mainframe problem. And it is this: "Some things do only work on SMP, because for some workloads no network fabric (that I know of) is as fast as the system bus." is under siege. From a systems architecture point of view it is a really interesting exploration of Amdahl's law. So many things that people said "You'll never do that on a networked cluster of machines." have fallen (data bases being one of the larger ones). And while it used to be mainframes won on I/O channel capacity, Google and others have shown that when you can parallelize the I/O channels effectively that advantage goes away as well.
- bogomipz 9y ago>"back in the early 2000's when "clusters" and "NUMA SMP machines" were competing with each other" Did you mean Numa vs SMP? Or something else maybe? How can a machine be a NUMA SMP? NUMA and SMP are fundamentally different architectures.
- zurn 9y agoIs this clear cut in the terminology? A plausible definition would also be that it's still symmetric MP if remote memory access has non-uniform performance - since the nodes and their memories are symmetrical. After all, you get that just with caches and 2 sockets plugged to the same DRAM. The historical opposite of SMP used to be asymmetric multiprocessors in the heterogenous sense - different kinds of processors, for example scalar/vector/io processors. Or for a modern day take, ARM SoCs with little low-power cores and faster & more power hungry cores, and GPUs thrown in for good measure.
- bogomipz 9y ago>"Is this clear cut in the terminology?" I think in the context that OP was referring to "early 2000s" which presumably means the introduction of Opteron and HyperTransport then yes I believe the NUMA vs SMP distinction would be pretty clear.
- ChuckMcM 9y agoNo, I was thinking non-uniform memory architecture, which is to say an SMP machine where the "speed" at which you can access RAM is dependent on the core or 'thread' from which you accessed it. Lack of memory uniformity was the compromise to achieve larger effective address spaces and "simple" programming. Today on a 24/48 core dual socket server you'll see the same sorts of thing, having a core using memory on the 'other' physical chip's memory bus will impact the overall performance significantly. In the 2000's, before multi-core chips were a thing, there were two major camps, the 'super computer' camp, and the 'cluster' camp. The 'super computer' camp insisted on cache coherent memory between all of the cores or threads. You got these very expensive fabrics from people like cray that would snoop access to memory from the cores and send coherency messages around to insure that if someone wrote something in to memory somewhere, everyone else's L1 or L2 cache got the message to invalidate what they were holding (shoot downs). These machines are very expensive and take months to build. The 'cluster' camp said, "We can use a network fabric and just parameterize shared memory usage." So they put together independent machines connected by a network fabric and no cache coherency protocol. If you wanted to use shared memory you could build something like memcached and wrap your access with network calls. With that architecture even if it took twice as many cores to do what you wanted to do, the price of the machine was one tenth what it was for the big SMP machine. For something that was trivially parallellizable like internet search or serving up web sites, that was a much more cost effective way to go. When people started doing stuff that they previously used 'super computers' and big SMP machines for on these Linux clusters it became a sort of race to pull apart these problems into "shared nothing" clusters.
- godelmachine 9y agoHi Chuck, by SMP do you mean Symmetric Multi Processing? Sincerely,
- ChuckMcM 9y agoYes. All threads have equal computation capability and cache coherent (although not uniform) access to main memory. This was the 'big iron' of the day like the Sun, SGI, and Cray "big iron" boxes.
- mathattack 9y agoGood to see you outside the context of HN. You're an immense value to the community.