Hacker News Reader

Comment by apalmer | Hacker News Reader

Ehhh, I am not sure that this is truly the issue with your presentation.

If it is truly "we can do X just like FamousApplication(tm), except we are in the browser and we can be profitable charging 10% as much as FamousApplication(tm)", that is pretty compelling. That is not something that is really difficult to sell if you can truly deliver and be profitable at that price point.

I am not sure how it is possible you can replace a computationally expensive client side application like MatLAB with something that runs in the cloud, without discovering order of magnitude more efficient solutions for a wide array of well know mathematical algorithms. If you have discovered such then you can pretty much write your own check, if not there is no way you can scale as a service. Thats what I would think you have to convince potential investors about.

View on HN

"I am not sure how it is possible you can replace a computationally expensive client side application like MatLAB with something that runs in the cloud, without discovering order of magnitude more efficient solutions for a wide array of well know mathematical algorithms. If you have discovered such then you can pretty much write your own check, if not there is no way you can scale as a service. Thats what I would think you have to convince potential investors about."

I am an engineer (not CE/CS) who uses Matlab, so I am probably missing something fundamental. However, I am curious why you think server side Matlab requires order of magnitude more efficient solutions to current Matlab. Even as a poor college student I used to SSH into my personal desktop running Matlab and it ran acceptably.

At work I will remote into another system running Matlab without issue. Sure if you are doing heavy plot manipulation it may get a bit laggy depending on the situation, but why would mathematical algorithms need to be rewritten unless you mean graphical?

I am basing it on the assumption that you cant pragmatically do 100 matlab sessions on a single server or VM. While a single Matlab session may run great on a server you are SSHed into, i dont think you can run 100 matlab sessions on a single VM or server without problems. Maybe you can in which case great!

If you cant though, then there is some kinda balancing act that the OP has to optimize, how many matlab sessions can simultaneously be hosted on one of the OPs VMs in the cloud at what cost? Since most of the cloud infrastructure providers seem to charge an arm and a leg for CPU, and I am assuming (maybe incorrectly) that this service will be CPU heavy, as an investor I would have to be convinced that this team has some technique for bringing down these costs...

either they would have to have better algorithms, or perhaps some interesting grid architecture that drastically reduces CPU time and hence costs, or something.

I am making assumptions here, but I think these are standard assumptions, these are the kind of assumptions you have to convince a investor you have a solution for. It may be as simple as saying, we can service 10,000 simultaneous users on 10 VMs, I have no clue what the numbers work out to, but they would have to at least show how this can scale profitably.

"I am basing it on the assumption that you cant pragmatically do 100 matlab sessions on a single server or VM. While a single Matlab session may run great on a server you are SSHed into, i dont think you can run 100 matlab sessions on a single VM or server without problems. Maybe you can in which case great!" "10,000 simultaneous users on 10 VMs"

Thank you for the response. I appreciate the insights but I do not agree with the premise that 100 or 1,000 Matlab sessions per server or VM is crucial to the economics of the business.

I know many engineering firms who are perfectly happy and profitable to buy a $2,000+ workstation, $2,000 Matlab license, and X*$Y,000 toolboxes for a single engineer. That engineer might be in meetings all day. They might be on vacation, they might not be working on a task that requires Matlab. Larger firms sometimes get floating licenses that may be 10X or more that cost and still have less 100% software utilization.

I see utilization economies both in software and hardware by licensing out this resource to small and mid-size firms who do not have constant large demand and even large firms who need a surge in computation.

Hey man, just wanted to put it out there, I am sure you have a much better understanding of the space than I do. I dont even think these are 'deal breakers' i just think they are questions that you have to be able to clearly answer when looking for investors... you may have covered them in depth during your interview, just seemed like your original article was trying to divine a mysterious missing puzzle piece that caused you to miss acceptance at ycombinator... from the little info we had to go on seemed to be some basic normal questions that weren't resolved before worrying about the philosophical something or solution arguement

I sincerely wish you great luck and success!

They could just spin up an EC2 instance when someone connects. It's $0.25/hr for a Large (~ 8gb ram, 2/4 core), so if they're servicing someone for a working day (8 hrs), they're out $2. If the user wants to save that working state, they can image the EC2 instance before turning it off (or reassigning it to a different user).

It would also be a nice feature that if you want to do some heavy calculation, you could spin up something more heavy-duty (a high-cpu xl is only 50c/hr). This same workflow with matlab would require you to go and buy faster hardware.

They could do all of that with some kind of standard matlab clone and probably get customers. If on top of that they had some parallelization secret sauce (some way to spin up new ec2 nodes on demand to do calculations), then they've really got quite a product.

The problem isn't the networking part.

The problem is running multiple CPU bound processes on a single server. Cloud systems are great for stuff requiring storage. It's super easy to make a 1 Tb disk look like one hundred 10 Gb disks. The way files are used and the way storage systems are designed, it's transparent to the user that their 10 Gb "disk" is really a VM's fake hard drive.

But it's not possible to do the same thing with the CPU when any of the users are running CPU bound computations. Performance would be terrible. It's particularly bad for Matlab/Octave because they can take advantage of multi-core CPUs. For a lot of stuff the problem could be alleviated by giving each user their own CPU core, but with Matlab it would mean their code is running 1/8th the speed they could get if they ran it locally.

looked at your product, it appears to be quite literally octave in the cloud.

Serious question, How can this scale to tens of thousands of users simultaneously using it? It is entirely possible that this is no where near as performance intensive as I am assuming it to be.

Second question, Whats to stop a competitor from doing the same thing your doing at half the price? It seems like it would be relatively trivial for another company to take input matlab programs from a browser, shell out to octave, and then return the results via the http response, and make it look pretty on the client. There may be a whole lot more going on, but from the outside looking in it appears like there is no moat, any technically decent company could enter this niche and do the same exact thing your doing with minimal effort.

Those two questions you have to answer i think rather than whether or not this is a 'solution' or a 'something'

The product right now looks pretty crude, but it also looks like a very early stage prototype.

If they had say 5 good engineers on this for a year, I suspect it would look very much more compelling.

To do that, you need funding, hence the ycombinator interview.

Reply on news.ycombinator.com