Comment by lmeyerov | Hacker News Reader

Comment by lmeyerov | Hacker News Reader

Fun to see a link here!

Always happy to describe what's happening underneath w/ connecting GPUs in the browser to GPUs in the datacenter. Likewise, the connection between event data & graph analytics is powerful as data scales, so happy to dig into that too.

Not shown there, we're piloting a 'visual playbook' investigation layer to help teams who investigate through a lot of event data. This has been especially relevant for security (SOC/IR/hunt) and anti-fraud as a team grows and needs to cover more ground. Playbooks let you finally record common multi-step multi-datasource workflows and get real visibility out of them. When your alerting flags something, running the playbook will gather & correlate that data for you, and unique to Graphistry, present it in a full visual (graph) analytics session. Think visually automating multi-step queries across Splunk + Spark + various APIs. And of course, for the advanced analysts, giant GPU-accelerated visualizations. We're actively piloting with interesting teams, so please ping info@graph....com if it may be your team's kind of thing.

Is it possible to visualize graphs entirely on the client side (without sending any data to your backend)? We have some very large graphs that we'd like to explore, but unfortunately it's not possible to send the data to the cloud, hence a local solution would be great. I have investigated Gephi but unfortunately the performance is quite disappointing for very large graphs.

Graphistry is impressive, but not free, and needs their server and an API key. You probably already know about graphviz sfdp (not interactive). There's been interesting recent work on distributed maxent optimization https://arxiv.org/pdf/1506.04383.pdf and sparse stress models https://arxiv.org/abs/1608.08909 (see its references as well, for example, Khoury's MARS is still available at https://github.com/marckhoury/mars). This problem requires a ton of algorithm engineering, a lot of messy user interface engineering, and integration into an HPC environment is a big issue, so progress in open source has not been that rapid.

Those projects are pretty exciting. We accelerated FA2 and are now adding acceleration to the surrounding analytics interactions. And yep, to support an eco-system here, we've been iterating on a sustainable & scalable model (... "enterprise").

To that end, we have a variety of non-profits & scientists using our cloud tier. We want to streamline that program in the coming months: If that's you today, we'd love to help, so feel free to reach out for early access. For example, we've been excited by data science and cybersecurity schools, journalists, and scientists.

Not sure how large very large is or the performance you need but you might want to check out the fairly recently added graph support in datashader (examples: https://anaconda.org/jbednar/edge_bundling/notebook)

Here's an example of the output which I then put into tiles & rendered out: https://proseandcode.co.uk/beta_gtr_viz/

I think there's 600k edges in this, all rendered.

I did some of the work on the edge bundling, the split out code is here: https://gitlab.com/ianjcalvert/edgehammer

All of this is more work than just loading it into a current program though, but it might be a useful component of what you need, hopefully a large useful part.

Get in touch if I can help out.

I'm a fan of the continuum team - we have been working together on the GPU GoAI project. We'll load 2M+ edges, and edge bundling helps make that visually useful.

Graphistry is a bit different where the result is a full interactive visual analytics session, not a zoomable png. So you get visual filtering, histogramming, search, etc. Our goal is to get from a lot of data to the answer, including whatever data pivots/cleaning/etc., as fast as possible, and that includes helping analysts skip a lot of the visualization/data coding and instead do direct visual interactions.

Worth stating: Edge bundling is beautiful! It was an early algorithm we implemented. A bit differently, we did it interactively, so you could do things like adjust sliders real-time to get the right physics settings. However, we found you'd want to hover over individual nodes/edges to see what's in them, and "zoomable image" style makes that hard. I've been wanting to bring it back now that we're getting close to supporting dynamic grouping interactions, so cool to see you call it out.

That looks impressive, will definitely check it out!

Indeed, most of our users run on-premise / air-gapped. I recommend trying our cloud API to make sure it's in the right ballpark, and then we can get you setup!

Is it possible to use a local GPU from the machine that has the data? Kind of like Gephi?

Yep! And we're working towards multi-GPU boxes!

Nice! Thank you

Reply on news.ycombinator.com