I like Cassandra and I think it's something you should look at hard if what you want is an eventually consistent database because multi-DC and availability are important to the problem you are solving.
The blog post fails to make that case anywhere. I am guessing that is because it is an ad for GCE and not Cassandra.
I know these kinds of benchmarks are necessary to get and keep your widget on the map, but they always annoy me because I know N other ways to arrive at the same or better numbers with a different set of tradeoffs. Demonstrating scale out is not great for differentiation because it says nothing about the total cost or complexity of use.
Why is it that only GCE + Cassandra could have done a million of these particular kind of writes in a cost effective way?
I also feel like they are fudging how elasticity works. It's not instant. Bringing up a cluster of a given size is not the same as adding node at time to a running cluster to arrive at that size and actually receiving full benefit from the additional capacity. From the wording it sounds like the entire cluster was brought up at once.
Overall my beef is that the blog post fails to inform the reader and is mediocre as a benchmark of anything other than GCE. And let's not even start with the complete lack of random reads, does GCE offer SSD instances yet?
I ran those tests. One of the challenges in benchmarking is to pick the workload, so I chose one our customers do more often. For IO testing I prefer to run FIO tests, and for networking iperf, but unless you do this for a living it is hard to relate to microbenchmarks. It would also leave out the Java settings.
The reason I left the tests running for an hour is because at some point Cassandra gets into compactions, which are CPU and read intensive. The engine memory maps the files, so the IO subsystem backing the storage sees a lot of random IOs as page faults kick in. Streaming IOs get in during fsyncs. Again, easier to see using FIO.
The cluster was brought up at once, and I added the ramp up time on the chart so folks could see it.
I had three goals for this test:
- Show low latency is possible with remote storage and proper capacity planning. This is why I used quorum commit, which forces the client to wait for at least two nodes to commit.
- Share the settings I used. If you open the GIST and download the tarball you'll find all changes I did to the cassandra.yaml and cassandra-env.sh. This benefits our customers directly because it gives them a starting point.
- Recommend that customers look at all samples, not only the middle 80% the stress tool reports. Otherwise the cluster looks much better than it is.
Thank you for the comments! I can assure you I read them, and I'll incorporate suggestions as possible.
Thanks I missed the link to the tarball in the GIST. Now that I can see the workload I see that it is 100 million keys for the entire cluster of 330 nodes?
That is 8 terabytes (n1-standard-8) of RAM for 18.6 * RF gigabytes of data or am I wrong? The entire thing should fit in the page cache assuming compaction keeps up. There are no deletes so no tombstones. On an overwrite workload I don't know what the space amplification will be and whether you will actually run out of RAM for caching.
90% of people who read a benchmark aren't going to look at it the way I do. They don't have a cost model that says how fast things should be and how much they should cost. I do and when things are fitting in memory I have a different a different set of expectations. If I am looking at the wrong instance type please let me know.
For the workload you described Cassandra shouldn't be doing random IO. I would expect there to be three + N streams of sequential IO. The write ahead log, memtable flushing, compaction output, and N streams reading tables for compaction.
All the write IO can be deferred and rescheduled heavily because fsyncs are infrequent. Reads for compaction are done by background tasks and shouldn't effect foreground latency.
If read ahead is not working for compaction (and killing disk throughput) that may be something that needs to be addressed. Compaction should be requesting IOs large enough to amortize the cost of seeking. Page faults in memory mapped files don't stop read ahead and I think that the kernel will even detect sequential access and double read ahead. For a workload like this with no random reads you could configure the kernel to read ahead 2-4 megabytes and the page cache would probably absorb it fine.
For the workload you described the only fsyncs would be the log (every 10 seconds) and memtable flushes/compaction finishing and that would normally be so infrequent as to not move the needle on overall IO capacity (although obviously it consumes sequential IO). You set trickle fsync so we are still talking about 10 megabyte writes.
Granted there are many things I don't know about Cassandra. I've plumbed a lot of it, but it isn't my day job. Using a 24 gigabyte heap and a 600 megabyte young generation is questionable to me. I think Cassandra can do better, and I also think there are several tools that would do the same job with 10-20x less nodes by exploiting the fact that they never have to do random reads from disk.
Hi - I figured you know what you're doing based on the details you picked on. I suspect you'd have way more fun reading the raw logs, but that would make for a post that is incredibly hard to parse. I will try to figure out a venue for them.
It was 100M keys per loader, so 30x100M for the low latency cluster and 45x100M for the high latency cluster. I sized to keep running long enough to go over half dozen major compactions, not necessarily to eat storage space.
Most of the IO is sequential, which really makes no difference for our backend (sorry, I have to leave it at that). Read performance ends up affecting the performance when the system reaches its limits of pending compactions, which I limited in the config. I could probably get a higher number without it, but it looked more realistic to set limits and I'd not deploy a server without the limits.
I had to make the fsyncs more frequent, not less, to reduce the odds of an operation sitting behind a long flush. This was a counter-intuitive finding for me, but it makes sense considering Persistent Disks throttle both IO sizes and IOPS. We do that to make sure a noisy neighbouring VM won't affect yours. So I turned on trickle_fsync and tuned the size of the maximum syncs to never have a slow flush, but a lot of smaller flushes.
Tuning the Java GC was for the same reason - I did not want long GC cycles. The settings I used were based on the guidance from Java memory ergonomics. The DataStax distribution has tips and hints in the cassandra-env.sh file itself, which I read it and followed through. I also read more about Java memory than I ever want to read again from various vendors. As the joke goes "I did even the things that contradicted the other things" before getting to the settings you saw.
I hope you appreciate the fact I set limits on some of the most dangerous knobs, for instance limiting the RPC server and using HSHA as opposed to unbounded sync.
I understand the limitations of the benchmark I used, and the limitations of the post. I tried to stay within recommended settings for all subsystems, have safe limits for everything I found dangerous, and I picked the workload because that is what our customers do. I don't advocate one storage solution over another, we love everyone that buys our solutions :)
> The cluster was brought up at once, and I added the ramp up time on the chart so folks could see it.
Can any person really bring up a cluster of this size? On AWS, you need special permission to launch more than ~20 instances of one type, and it's granted only after you make a business case for it, which they won't grant to regular peons. With my regular Google account that doesn't have any kind of special approval, can I really launch 330 instances within a few minutes?
There is a CPU quota limit (and a disk limit) that you have to get approved to raise (and I think there's a limit to how many requests/second you can make to the GCE API), but if you get those approvals, you're good.
In the end these are just eye-popping benchmarks. What really drives people to your platform is developer friendliness and the speed by which you can get started. Google App Engine has this, but there's not a lot of information floating around the media and blogosphere as there is for Heroku and DigitalOcean.
I use Google App Engine for auto scaling an API and it works brilliantly. Super easy to set up and develop on. But recently I started preferring DigitalOcean simply because there is a community that is constantly posting tutorials and answering questions. To me, that's more valuable than the distant prospect of handling 1M writes/second.
Looks like a great way for DO to both build a sense community and also appear high in Google rankings when you search for something. I googled some Linux question earlier today, and ended up on a DO page -- refreshing their brand in my mind.
While this doesn't directly pertain to your comment, perhaps it's somewhat related to onboarding with GAE...
I've been developing on GAE since it was released in 2008 and recently became "Google Cloud Platform Certified".
I'm considering offering a "Google App Engine for Startups" class/workshop (at something like General Assembly). The class would primarily include details about the platform's architecture and best practices for building high-scalability apps (like, say, Snapchat) so your app can "scale without thinking twice." Do you think there'd be some interest in a class like this?
From organizing past events, it's much easier to just do it and have it flop (nobody's going to notice it if it does) than try looking if there's interest or not. Ask your target venue if you can do it, make sure it has some visibility (ie. make sure it's in the program), setup a slide deck or two and do it. Good luck!
To the point about what would attract users to a platform, offering a developer friendly environment and quick ramp up time, I certainly agree.
The 1M writes/second would be a capability more specifically focused for users/developers interested in using already existing platforms such AWS's High Performance Computing platforms and the like.
The "one million" number grabs some attention, but this isn't that special in my opinion. Doing 3333 writes/sec per node is not that hard with Cassandra, actually it can be much faster if you use a setup with fast local storage and split for example the data disk from the commit log. The Google article reads as if they used network storage and 1 volume per node, both are bad ideas for Cassandra as documented by Datastax.
Keep in mind, it's not "One Million Writes Per Second," it's "One Million Writes Per Second on Google Compute Engine" with "Google Compute Engine" being the key point to the article.
It is worth noting GCE is more expensive now that AWS was back in 2011.
According to Netflix article the AWS experiment did run at a cost $561 per 2h, that is ~$280 per hour. Perhaps they were not utilized the cluster fully in those 2h in which case we should multiply the 1h test that performed 500k inserts per second, in that case the cost would be $182*2 = ~$365 per 1h.
GCE test did run at the cost of $330 per hour. Give or take few dollars difference if anything it's surprising GCE can do at roughly the same cost what AWS was capable of 2+ years ago.
Saying all that GCE guys did a great effort. I wonder though how much speed you can squeeze from AWS and at what cost now when AWS is sporting SSD disks.
Hi,
The cost we published includes the time to setup the whole cluster, warm up the data nodes, and run for 5 minutes at 1M per second.
Our run rate is $281 per hour, which is the same as AWS a couple of years back. What changed is that we are using quorum commit, the data is encrypted at rest, we have very low tail latency, and we look at all samples when computing that.
Computing our price is easier because we do not charge per access.
Here is the formula for our run rate:
30 loaders (n1-highcpu-8) at $0.522 per hour: 248.7
300 nodes (n1-standard-8) at $0.829 per hour: 15.66
300 1TB PDs that run at 0.055555556/hour: 16.67
Total: 281.03
I wrote the test - Yep. Tail latency is one of the key things here. And I took 100% of all samples, as opposed to the middle 80% the tool usually reports.
Network average utilization was low by design. Keeping it steady was more important than low, though, and harder too.
Latency spikes come from Cassandra flushing data to disk (large sequential IO), Java garbage collection and heap resize, and page faults during compactions (random reads).
What I did to even traffic out was to enable trickle_fsync and size the flushes, set Java's max and min heap sizes, as well as to tune the Java heap ergonomics. I treated random reads as a fact of life - I did nothing to tune that.
Doesn't GCE run on the same (physical, not logical) network as the rest of Google's production systems? If so, which I believe is the case, how can you control for network utilization?
Now I would like to see how many reads it can do at the same time. Cassandra does mostly-sequential writes, however most reads do couples of disk seeks.
Can anyone comment on Cassandra's performance versus HBase? HBase adds the complexity of dealing with the whole HStack and clustering can be a pain if you have no need to use HDFS/Zookeeper in the first place. Cassandra seems nice because its a single platform and each node is an equal member of the cluster, no need for designating HMaster, data nodes, etc. I was just wondering if Cassandra's widely regarded as less performant.
I know the true answer lies in my exact use cases and weeks of initial testing, but it would be nice to hear someone's opinion first.
Is anyone actually using Google Compute Engine? I haven't bumped into anyone using it and would love to hear what the real world experience is with it.
I've tested it, network latency is by far not as good as for example Rackspace Cloud. And tool-support is very far behind Amazon and OpenStack. Basically Google Compute Engine is something that might have had a chance years ago but at this point in time it's quite pointless because it has no upsides compared to existing far better established solutions and is behind on features, documentation, tools and user-support.
And I expect that not to get better soon if nobody is using it...
> And I expect that not to get better soon if nobody is using it...
If the people running GCE see that people are not using it because it sucks at X, they will probably do their best to address X. So complaining is a good thing, especially if it contains enough information in order to understand the problem (e.g. latency from/to where? internal network or external network?)
> And I expect that not to get better soon if nobody is using it...
If Compute Engine is Google exposing internal infrastructure as an additional product, its quite likely that lots of people are using it (and that it is strategically important for it to be more usable) even if few people are paying for it.
The 1M writes/second capability is a valuable benchmark in analyzing continuously variable and real-time information, similar to CFD analysis are high-throughput computational analysis where you end up dealing with multi-variate dimensions of data continuously in flux.
Unix admins have been using a far better technology for decades: Store in /dev/null, extract from dev/random. The write speeds are just as fast, but you will eventually get your results back. It is said that it'd be quicker to extract the next Game of Thrones novel from this system than to wait for George R R Martin to finish writing it.
1. What is the point of benchmarking so many writes? Instead, give us the cost of write-read and a few use cases: analysis, data mining, etc. And what of the bandwidth costs?
2. Say you are running a moderately intensive application that needs to be up 10 hours a day. $330/hour x 10 = $3300.
Multiply that by a few days and where does that leave you?
Suddenly having your own infrastructure with kit like this:
http://www.fusionio.com/products in it is not such a bad idea! This also means you don't have to use any funny NoSQL products, can use standard programming, non cloudy infrastructure and a decent RDBMs setup that non Silicon-valley, non-hipster humans have a chance of understanding and reliably maintaining.
The blog post fails to make that case anywhere. I am guessing that is because it is an ad for GCE and not Cassandra.
I know these kinds of benchmarks are necessary to get and keep your widget on the map, but they always annoy me because I know N other ways to arrive at the same or better numbers with a different set of tradeoffs. Demonstrating scale out is not great for differentiation because it says nothing about the total cost or complexity of use.
Why is it that only GCE + Cassandra could have done a million of these particular kind of writes in a cost effective way?
I also feel like they are fudging how elasticity works. It's not instant. Bringing up a cluster of a given size is not the same as adding node at time to a running cluster to arrive at that size and actually receiving full benefit from the additional capacity. From the wording it sounds like the entire cluster was brought up at once.
Overall my beef is that the blog post fails to inform the reader and is mediocre as a benchmark of anything other than GCE. And let's not even start with the complete lack of random reads, does GCE offer SSD instances yet?