# Notes on using Jupyter in the cloud

## Notes on using Jupyter in the cloud

I've been thinking about running Jupyter notebooks in the cloud for some fairly compute-intensive simulation. Specifically I want to do epidemic and other simulations over complex networks. These are CPU-intensive and don't make use of GPU acceleration (yet, anyway). Using the cloud would make things easier to scale-out, especially for those without access to a local compute cluster.

I decided to do some experiments by writing a small Jupyter notebook that would exercise a cloud service by running some simulations using epydemic, our network process simulation framework. The idea was to see how performant the cloud services on offer are, starting with Google Colab as the most accessible.

### Reference execution

As a reference, I firsat ran a small experiment with an SIR epidemic on an ER network of $10^4$ nodes, $\langle k \rangle = 10$) experiment with 10 repetitions. The first experiment runs 10 repetitions sequentially on a single core; the second runs 10 repetitions for every core we run on. epyc, our process management library, schedules worker processes in individual processes which can then be scheduled independently by the operating system, and which don't therefore run into synchronisation problems due to the Python Global Interpreter Lock.

Running this on my desktop machine (Intel Core i7@3.8GHz) on 8 cores and taking averages gives:

Cores Elapsed time (s) Mean/rep (s) Variance
1 38 3.86 0.06
8 43 4.27 0.09

There's some overhead from the collection of the experimental results from the worker processes.

### Running epydemic on a default runtime

I then tried the same experiments on the default Colab set-up.

The notebook can import libraries no problem. It can also create HDF5 notebooks, meaning it has the underlying libraries installed already. The GUI reports 12GB of RAM and 100GB of disc (approximately).

We run on all the available cores, as reported by the epyc.ParallelLab.numberOfCores() method. Running the same experiments as above gives:

vCPUs Elapsed time (s) Mean/rep (s) Variance
1 81 8.17 0.90
2 100 10.87 0.34

which is a little unusual: there's definitely some overlapped computation happening, but with large overheads and large variance. I'm assuming that the number of cores reported is actually the number of hyperthreads, and that they're interfering because they're CPU-bound: epyc simply uses the standard Python joblib. Or it might be that the underlying VM isn't really allocating cores exclusively. There's really not enough information to decide.

### Using specific VM instances

I signed up for Google Cloud's free trial, 90 days and $300 (£281) of credit. Then set up a specific project in which to run the experiments, and explored the different VM instance types available. E2 and N2 are the general-purpose instances, available in standard or high-CPU configurations. The E2 is cheaper than the N2 but less performant – although this only cuts in at higher loads according to this comparison. Compute-optimised VMs (C2 and C2D) are Intel Xeon or AMD EPYC Milan processors. The C2D has a larger last-level cache and so gets recommended for HPC workloads. The VM instance pricing model charges for a single minute, and then after the first minute charges per-second. Instance vCPUs Memory (GB) On-demand price/hour Spot price/hour e2-highcpu-16 16 16$0.395744 $0.11872 e2-highcpu-32 32 32$0.791488 $0.23744 n2-highcpu-64 64 64$2.294272 $0.55552 n2-highcpu-96 96 96$3.441408 $0.83328 c2-standard-60 60 240$3.1321 $0.28428 c2d-standard-112 112 448$5.0844 \$1.077776

That's quite a range of possible costs.

The Spot VMs are cheaper, but make use of spare capacity and so might be pre-empted (or terminated) by higher-priority jobs. They're recommended for fault-tolerant workloads, which we could possibly get if we extended epyc with some extra back-end functions. You also can't buy them with the free tier's credits.

There is apparently a limit of 8 vCPUs when using the free tier for some instance types in some zones. It's not abundantly clear what the restrictions actually are.

The process to create the custom runtime is to create a Colab VM instance from the Marketplace and associate it with an instance type of choice, possibly choosing a zone as well. This can then be connected to by providing the project, zone, and instance name to the Jupyter notebook. (The project name has to be in URL form, so lower case with dashes for spaces.) This is awkward and requires manual copying, but once done you can acquire a connection URL to go straight to the notebook running on that VM. The Jupyter UI shows RAM and disc as well as the actual VM instance it's connected to, but not the number of cores that instance reports, which is a bit annoying.

There's also an issue in that, when you ask the machine for the number of cores it has, it by default replies with the number of vCPUs – which I think means hyperthreads. A 96-vCPU machine (an instance ending in "-96") only has 48 cores, because by default 2 vCPUs are mapped to each physical core. You can set the ratio of vCPUs to cores (1 or 2), and the numnber of visible cores the machine reports. So I set a ratio of 1 vCPU/core and reporting the number of actual cores, which is the sensible choice for a compute-bound application. Unfortunately you can't do this without stopping the newly-created Colab VM and re-setting it's configuration: you can't do this step at instance creation from the Marketplace. I don't know why. (It might be possible to do it in one step from the command line. Or create template instances with the right configuration.) On the other hand, once it's done, it's persistent and can be connected to using the connection URL, as the notebook remembers the VM it's connected to.

### Experiments on specific instances

Running the same experiments as above on different instances gives:

Instance Cores Elapsed time (s) Mean/rep (s) Variance
e2-highcpu-16 1 98 9.81 0.09
8 96 9.49 0.10
n2-standard-8 1 74 7.44 0.23
4 69 6.83 0.08
c2-standard-8 1 70 7.05 0.05
4 66 6.55 0.07

(These are real cores, 1 vCPU/core.)

There's around a 30% speed difference between the E2 and N2 silicon, but not much at all between the N2 and C2 – despite the latter being branded for compute-intensive workloads. Might be that the C2's cache isn't being exploited?

We do however get the speed-up we expect from parallelism: actually slightly more than we'd expect, since the individual runs seem to go faster too. There's definitely some overhead incurred in running epyc in parallel, so we shouldn't see super-linear speed-up "in reality".

### Running larger problems

For more of a soak test, we can run the same SIR experiment but using a larger ER network ($10^5$ nodes, $\langle k \rangle = 10$):

Instance Cores Elapsed time (s) Mean/rep (s) Variance
e2-highcpu-16 1 1235 123.56 1.21
8 1216 120.30 0.86
n2-standard-8 1 930 93.09 0.56
4 869 86.51 1.06
c2-standard-8 1 911 91.18 0.72
4 849 84.53 0.41

There's that super-linear speed-up between sequential and parallel versions again.

The performance on the standard runtime for comparison is:

vCPUs Elapsed time (s) Mean/rep (s) Variance
1 1070 107.60 3.84
2 1402 140.22 1.50

### Costs

Doing all the above experiments used rather less than £10 of the budget for my free trial – although I was very careful not to leave instances running when I wasn't actually using them. This is an unusual thing to be considering, not part of my "normal" work routine, and would possibly be awkward for longer-running computations. You'd be reluctant to run something overnight if you weren't sure it needed all night, for example. This might be addressed by using the command-line tools to spin-up, execute, and then tear-down the infrastructure using a script.

### Experiences: Good and not-so-good

Good:

• Everything controlled from a web console
• Easy to run pip to install dependencies
• Once installed, the dependencies persist even if the VM is shut down
• The GUI shows how long cells take to execute, as well as the memory and disc of the underlying machine and its instance name
• There's a set of command-line tools
• Persistent links to notebooks
• Notebook remembers its connection to the underlying VM instance

Problematic:

• All the available instances are considerably slower than a reasonably modern desktop workstation
• If an application needs more than just pip dependencies, that'd have to be done at the VM level using ssh etc
• Fiddly sequence to get vCPU and core reporting appropriate for HPC
• Need to manage spin-up and tear-down of instances, and incur costs if you forget
• The GUI doesn't show how many cores the underlying instance has
• The management console requires a fairly decent knowledge of cloud computing concepts, which need to be learned somehow. I'm not convinced the tutorials on the web site are good enough for someone without plenty of background
• The notebook doesn't seem to deal cleanly with disconnections, which is a problem if you have a flaky connection