Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- curious how folks like to measure this stuff wrt load testing?
We are actively revisiting our traffic simulation approach, and a surprisingly non-obvious part has been which charts to focus on. Our case is a gpu-server-backed interactive analytics app, like a photoshop for data, so we do focus both on latency and errors, and especially around handling bursty sessions as discussed in the article.
by lmeyerov - The problem with this kind of theoretical analysis is that most load balancers don't work this way, especially the typical "cloud" HTTP or TCP load balancers, which are stateless and avoid this kind of central queuing logic like the plague because it doesn't scale to their levels.
For example, most cloud load balancers I've worked with are stateless, non-queuing, and allocate work to back-ends strictly randomly.
Traditional non-cloud load balancers can implement this kind of perfect queuing, but these settings are generally off by default even when available.
- NetScaler: surgeProtection + maxClient=1
- F5 BIG-IP LTM: request queuing + pool/member connectionLimit=1
- HAProxy: server maxconn 1 + timeout queue
- NGINX Plus: server max_conns=1 + queue
Envoy, Apache, and Traefik have partial or limited support.
Conversely, most multi-threaded web server frameworks already do this by default! For example, ASP.NET has essentially an internal "load balancer" with a perfect queue if you pretend each core is a "node" and the whole server is the "scale out system".
by jiggawatts - > can only handle a single concurrent request, and has no internal queuing
And the systems that load balancers front almost never behave this way...
Dont even get me started on client performance here as well, latency, speed, caching -- that can all be impacted by payload size.
The article is interesting, but it is an ideal that almost never turns up in the real world.
by zer00eyz - The footnote on exponential vs. log-normal service times is the part I'd push on.. in production I almost never see exponential, and heavy tails change the picture. Curious if you've looked at how robust this is under realistic distributionsby fabijanbajo
- It's not surprising if one has the mental model of the probability that the request gets enqueued. Then when you add variable time to process requests it becomes more clear why some requests can take unexpectedly long (there is a >0 probability that a request gets queued behind several of the slowest endpoints, for example). So even if 90% of the endpoints are fast and most of the requests aren't even queued, there will still be some that end up being quite slow.by resters
- Why would anyone think that it would get linearly worse? What's the (wrong) assumption there?by nilsherzig
- I think author made it up just to have something more to show up on graph.
- I think people are reading it as a request every 0.8s that takes 1s to process, instead of 0.8 requests per second.by gm678
- I thought the same thing. But, should we be surprised about what people believe in these days?
I think that the issue is in part due to the variables. Plotting the mean request time is less intuitive than plotting throughput.
If you plot throughput vs number of servers, it'll be a straight line. And asking people that, I think most would agree on a straight line. But who knows!
by physix - I'm mostly just surprised the graph starts at 5 seconds for a mean value for all datapoints. I would have assumed it starts much closer to 1s. Which just makes the poll responses even crazier. Who is picking B when you have 25% more capacity than you need?
But I suppose the question is underspecified. How does the load balancer know which systems are busy? What happens to a request if the load balancer routes a request to a busy server?
by jldugger - One explanation would be that more load could mean higher (absolute) variance in queue length, and therefore higher latency especially at higher percentiles. It doesn't work out that way (for reasons that Erlang actually writes about in one of his original works), but it's not an entirely unreasonable intuition.by mjb
- Of course, this assumes independent events. World Cup, super bowls, etc break these assumptions.
Still, queuing theory is so cool.
by crypttales - M/M/c is not the right model for a typical loadbalancer, since a loadbalancer typically does not manage a shared queue but simply passes the jobs to one of the servers. The models are:
- M/M/1 (vertical scaling, one queue and one fast server): fastest response time
- M/M/c (thread-pool, one shared queue, c slow servers): c-times slower for low loads, asymptotically similar to M/M/1 for high loads
- c-times M/M/1 (loadbalancer, c slow servers each which its own queue): always c-times slower than M/M/1.
Only the response times are different. Throughtput is the same in the ideal case.
by juergn - What's conspicuously missing is the plot of performance when you do have a well tuned queue in front of the service. Yes, having a queue becomes less important the more backend servers you have, but here even with 10 servers the plot shows your latency remains >25% worse than it would be with a queue. Also missing is discussion of how the variance in processing times affects you when you rely on load balancing alone.by megamalloc
- > What's conspicuously missing is the plot of performance when you do have a well tuned queue in front of the service.
As in between the service and the load balancer? There's already an infinite queue in the load balancer. You can try that out on https://stability-sim.systems/ to see the effect, but the short version is that (in this model) it makes things worse.
If you're saying that the queue in the load balancer should be limited in size to reduce tail latency, then I agree.
by mjb - A dead comment says:
> Of course, this assumes independent events. World Cup, super bowls, etc break these assumptions.
Yes, this is very true. The model here works for Poisson arrivals and exponential service time (the M/M), which are poor approximations of real-world traffic patterns (which tend to be non-stationary and non-ergodic, and include substantial seasonality). However, the frequency of that seasonality is typically rather low (e.g. daily cycles), and so these stronger assumptions are quite defensible for short time periods.
A better approach is to do simulation with real traffic patterns, or even with more sophisticated parametric models, and get better answers (e.g. https://stability-sim.systems/). The good news is that kind of simulation is cheaper to do than ever before.
by mjb - the article offers a simplified world model: Poisson arrivals and infinite queue, which is fine as a math model.
In the real world however, the bursts can be correlated, due to factors like timeouts/retries, thundering herd, correlated bursts.
so the real economics of load-balanced system is a simple reliability story: being able to reasonably serve the peak traffic, which leads to over-provisioning of those systems.
using cloud allows some form of scale up/down of resources, but doesn't completely solve the problem. I think the migration away from synchronyous systems towards async systems and letting clients gradually absorb the delays is a better approach (rather than forcing infrastructure to be dynamically scaled up/down and be billed per request-second by your cloud provider)
by bijowo1676 - >In the real world however, the bursts can be correlated
Very true, as application-layer load-balancing often explicitly pre-bakes the traffic schedule to several hundred distributed IPs for data locality. Essentially bypassing the functional need for DNS and local round-robin traffic balancers.
One trades concurrent bandwidth for slightly higher latency, and dynamically adapted capacity as traffic load changes. =3
by Joel_Mckay - > the bursts can be correlated
Hawkes processes are what other fields use to model this
- As in the classic paper "Wide-area traffic: The failure of Poisson modeling" by Vern Paxson (author of GNU flex) and Sally Floyd (legend in the world of TCP/IP congestion control):by fmajid
- Another technique to add to the mix if you can handle the additional complexity is to load or feature shed. If you can delay or just drop additional expensive application features during the exact time you need to scale or handle a burst, then your system has additional core app logic to handle requests. This can prevent the system getting wedged in a positive feedback loop.
See also the gamedev technique of having sacrificial assets or code, so when you need to free up space late in the schedule to ship, you have something you can actually shed.
by genxy - I'll argue though there is a political and ideological dimension to queuing theory, especially when it comes to systems where humans are handling the work.
In the idealized case the comfortable place to operate a system is with around 2/3 utilization, like around there latency (customer experience, employee experience) is reasonable, slack is reasonable, etc.
A manager who is being managed by a manager who is being managed by a manager who is being managed (...) is going to see 0.99 utilization and want that last 0.01 and be oblivious to the fact that the math says the system is already past the breaking point, customers are furious, employees are worn out. Any slack at at all seems like an affront.
by PaulHoule