I know SageMaker endpoints have autoscaling as an option, but from my understanding that mainly applies when there is a sustained high request volume. We have the issue that on occasion there will be a huge sudden single spike and then go back to normal. Is autoscaling fast enough (what's the delay?) for it to handle that? Or does it need to actually spin up another instance? Would having two instances at the endpoint help it respond immediately to an isolated request spike? I'm just not clear on what the response delay is for autoscaling, and I have do not see this mentioned in their posts/documentation. Thanks
for your question about time sagemaker need to start new instance is that the time can vary depending on the model size, how long it takes to download the model, and the start-up time of the container.
now what options you have :
use hw utilization let say when reach 76% of cpu scale out
another option to use step scale and for example use OverheadLatency
to scale out
here is a good resource for the above https://aws.amazon.com/blogs/machine-learning/configuring-autoscaling-inference-endpoints-in-amazon-sagemaker/
and you can always do load testing before you choice the best strategy that fit your need , check this url for load testing
https://aws.amazon.com/blogs/machine-learning/load-test-and-optimize-an-amazon-sagemaker-endpoint-using-automatic-scaling/
also , i think using aws sagemaker serverless is good solution for your case but it is in preview stage, right now
https://aws.amazon.com/about-aws/whats-new/2021/12/amazon-sagemaker-serverless-inference/
Related
We’re using TorchServe on SageMaker with the deep learning containers ( https://github.com/aws/deep-learning-containers/blob/master/available_images.md).
Everything seems to be running well, we have very minimal CPU and memory usage, and our latency of requests can be about 3 seconds (which is quite long but okay). Once we start to ramp up usage our model latency increases dramatically - taking up to 1 minute to respond, however, our CPU and memory usage is still minimal... we feel like the TorchServe instance is not scaling properly on the instance and only deploying one worker when it should be at least 4 on a 4vcpu system (ml.m5xlarge instance). We have been following official AWS guides.
How can we resolve this?
We work with large files (1-20Mb images). However, SageMaker has a limitation of 5Mb payload size. To get around this we are using S3 to store our intermediary files (which potentially causes long latency). Our images go through preprocessing, enhancing, segmentation, and postprocessing, which are currently all individual SageMaker endpoints that a lambda function controls. We want to use SageMaker pipelines, but we see no incentive to make changes to our system with the payload limit.
Is there a better way of doing this?
We have been experimenting with Multi-Model instances. Ideally, we would like to have an ml.g4dnxlarge run all our PyTorch models to benefit from the low latency and high throughput. Still, when we deploy a torch serve multimode endpoint with deep learning containers, we get an error when invoking "model version not selected" ???? I have looked everywhere for anyone having similar issues but can't see anything.
Are there any guides for torch serve multi-model instances? Are we making an obvious mistake?
We have experimented with elastic inference and AWS Inf instances. However, we have seen no latency decrease from the standard CPU instance - again,
are we making a mistake here, and the benefits from these instances seem beneficial to us, and we would love to use them?
App Engine has been great for requests that process quickly with no external API calls to databases or caches or third-party resources, but we've found that introducing any sort of "longer running" component or external latency (for example in a HTTP POST operation that runs asynchronously in the background and might take a second or two to process a few more intense database queries... totally invisible and OK from a UX perspective on the client-side because it's asynchronous but expensive to App Engine billing since it's long running) ... the "instance hours" compound and drive costs up considerably.
These sorts of expense inducing situations where a request is literally just waiting for a response from an external resource and requiring almost zero CPU during their idling seem avoidable, but I'm not sure if it's avoidable with App Engine.
It's almost like a "long poll" where the response might be left open but doing nothing.
Is there a way to do this on App Engine without just paying an insane amount for instance hours, or would we be better off moving to Compute Engine or EC2? Does it scale automatically based on CPU load, or is it based solely on open and perhaps inactive requests in total count? — threadsafe is indeed enabled.
There are really two ways to go about this one (top of mind).
Use Task Queues!
If the work doesn't need to be exactly at the same time of the request, this is exactly what [task queues] in App Engine are for. They allow you to put a job on a queue, and have another module pick up the work. They're kind of great because you can separately scale your front end and back end processes.
If that doesn't work....
Use App Engine Flexible
Under the hood App Engine Flexible is just running GCE instances. The cost structure is entirely different, since you persistently have a VM running in the background serving your requests.
Hope this helps!
What you're really worried about here is how App Engine scales your instances. Because many of your requests require few resources, your app might be able to handle many more concurrent requests on a single instance than normal. You can look into parameters that shape scaling here. Of particular interest:
max_concurrent_requests The number of concurrent requests an automatic scaling instance can accept before the scheduler spawns a new instance (Default: 8, Maximum: 80).
There is a danger here, where an instance may fill up with non-long-polling requests and become overburdened. To prevent that, you could isolate your long-polling requests into their own service and set its scaling parameters separately from the rest of your app.
I developed an application for client that uses Play framework 1.x and runs on GAE. The app works great, but sometimes is crazy slow. It takes around 30 seconds to load simple page but sometimes it runs faster - no code change whatsoever.
Are there any way to identify why it's running slow? I tried to contact support but I couldnt find any telephone number or email. Also there is no response on official google group.
How would you approach this problem? Currently my customer is very angry because of slow loading time, but switching to other provider is last option at the moment.
Use GAE Appstats to profile your remote procedure calls. All of the RPCs are slow (Google Cloud Storage, Google Cloud SQL, ...), so if you can reduce the amount of RPCs or can use some caching datastructures, use them -> your application will be much faster. But you can see with appstats which parts are slow and if they need attention :) .
For example, I've created a Google Cloud Storage cache for my application and decreased execution time from 2 minutes to under 30 seconds. The RPCs are a bottleneck in the GAE.
Google does not usually provide a contact support for a lot of services. The issue described about google app engine slowness is probably caused by a cold start. Google app engine front-end instances sleep after about 15 minutes. You could write a cron job to ping instances every 14 minutes to keep the nodes up.
Combining some answers and adding a few things to check:
Debug using app stats. Look for "staircase" situations and RPC calls. Maybe something in your app is triggering RPC calls at certain points that don't happen in your logic all the time.
Tweak your instance settings. Add some permanent/resident instances and see if that makes a difference. If you are spinning up new instances, things will be slow, for probably around the time frame (30 seconds or more) you describe. It will seem random. It's not just how many instances, but what combinations of the sliders you are using (you can actually hurt yourself with too little/many).
Look at your app itself. Are you doing lots of memory allocations in the JVM? Allocating/freeing memory is inherently a slow operation and can cause freezes. Are you sure your freezing is not a JVM issue? Try replicating the problem locally and tweak the JVM xmx and xms settings and see if you find similar behavior. Also profile your application locally for memory/performance issues. You can cut down on allocations using pooling, DI containers, etc.
Are you running any sort of cron jobs/processing on your front-end servers? Try to move as much as you can to background tasks such as sending emails. The intervals may seem random, but it can be a result of things happening depending on your job settings. 9 am every day may not mean what you think depending on the cron/task options. A corollary - move things to back-end servers and pull queues.
It's tough to give you a good answer without more information. The best someone here can do is give you a starting point, which pretty much every answer here already has.
By making at least one instance permanent, you get a great improvement in the first use. It takes about 15 sec. to load the application in the instance, which is why you experience long request times, when nobody has been using the application for a while
Lately, I have seen GAE taking much, much longer to process requests than it did just a week ago. Nothing changed in my code, but GAE now is taking 4000-12000ms to respond to requests. What makes is worse is that I have plenty of instances available with 0 requests on them.
Has anyone else seen this happen?
What can I do to fix it?I have gone as far as to spin up 15 extra instances (and paid through the nose for them) but nothing seems to send requests to the other idle instances reliably.
My bill has gone from 70-90c/day to $5-8/day without any code change or increase in traffic. In fact, I am losing traffic because of the huge latency.
QPS* Latency* Requests Errors Age Memory Availability
0.000 0.0 ms 1378 0 10:10:09 57.9 MBytes Dynamic
0.000 0.0 ms 1681 0 15:39:57 57.2 MBytes Dynamic
0.017 9687.0 ms 886 0 10:19:10 56.7 MBytes Dynamic
I recommend installing AppStats to get a picture of what's taking so long in each request. I'd guess that you're having some contention issues or large numbers of reads/writes caused by some new data configuration.
The idle instances won't help decrease latency - it looks like every request takes a long time, and with less than one request per minute (in this sample anyway), 10s requests could run serially on the same instance.
We have a similar problem in our app. In our case, we are under the impression that GAE's scheduler did a poor job in balancing requests to existing instances.
In some cases, the scheduler decided to spin up new instances instead of re-using already existing ones. Since spinning a new instance took 5 to >45 seconds, I suspect this might be what happened to you.
Try to investigate the following and see if it helps you:
Make sure your app has thread-safe enabled so that you could process concurrent requests. You could configure this in your app.yaml if you are using Python, or in your appengine-web.xml if you use Java. Of course, you also need to make sure that the code in your app is threadsafe.
In your application settings, if it is still set on automatic, change the minimum pending latency to a non-automatic setting. I'd suggest around 10 seconds for now, but you could experiment later on which setting would suit you the most. This force the scheduler to wait for a certain time to see if any instance is available within the time before spinning up a new instance.
Now, to answer your original question regarding sending all requests to same instance, as far as I know there is no way to address a specific front-end instance in order to direct the requests to that particular instance.
What you could do is migrate your app to use backend instances instead of the regular frontend instance. Backends provides a way to directly target any particular instance within it. You could deploy your app in a single backend to have more control on the number of instance that you spawn. And since using the backend bypass the scheduler, you would not encounter latencies caused by new instances spinning up.
The major drawback of using this approach is that you lose the auto-scalability benefit of using front-end instances. But seeing from your low daily billing, I think scalability is not yet a major concern for the scale of your app.
I'm using the Python 2.5 runtime on Google App Engine. Needless to say I'm a bit worried about the new costs so I want to get a better idea of what kind of traffic volume I will experience.
If 10 users simultaneously access my application at myapplication.appspot.com, will that spawn 10 instances?
If no, how many users in an instance? Is it even measured that way?
I've already looked at http://code.google.com/appengine/docs/adminconsole/instances.html but I just wanted to make sure that my interpretation is correct.
"Users" is a fairly meaningless term from an HTTP point of view. What's important is how many requests you can serve in a given time interval. This depends primarily on how long your app takes to serve a given request. Obviously, if it takes 200 milliseconds for you to serve a request, then one instance can serve at most 5 requests per second.
When a request is handled by App Engine, it is added to a queue. Any time an instance is available to do work, it takes the oldest item from the queue and serves that request. If the time that a request has been waiting in the queue ('pending latency') is more than the threshold you set in your admin console, the scheduler will start up another instance and start sending requests to it.
This is grossly simplified, obviously, but gives you a broad idea how the scheduler works.
First, no.
An instance per user is unreasonable and doesn't happen.
So you're asking how does my app scale to more instances? Depends on the load.
If you have much much requests per second then you'll get (automatically) another instance so the load is distributed.
That's the core idea behind App Engine.