Large-scale, worldwide scientific initiatives, such as the one that found the Higgs Boson or the one that is currently researching the depths of proteomics, rely on some cloud-based system to both coordinate efforts and manage computational efforts at peak times that cannot be contained within the combined in-house HPC resources.
Last week at Google I/O, Brookhaven National Lab’s Sergey Panitkin discussed the role of the Google Compute Engine in providing computational support to ATLAS, a detector of high-energy particles at the Large Hadron Collider (LHC).
On July 4 of last year, one of the largest physics experiments in history announced the finding of the Higgs Boson. The discovery was another step in the verification of the Standard Model of elementary particles, and it was largely a result of the data collected by the ATLAS detector that was later stored, analyzed, and used in simulations in computational centers around the world.
Naturally, CERN is equipped with significant computational capabilities as it sifts through the swaths of data created by the LHC. However, a great deal of that data was being sent out to scientists across the world in over a hundred computing centers located in over 40 countries.
As a result, Google stepped forward in August of last year to offer its Compute Engine services for overflow scientific computing periods. According to Panitkin, those spikes would occur before major conferences, overloading the existing computational framework. These overflow spikes represent an intriguing phenomenon, a macro-scale example of a problem that many mid-sized research institutions face on their own. Many of those institutions house their own HPC cluster that handles the majority of their heavy duty computational leg-work. When those resources are exhausted at peak times, they turn to the cloud.
When that problem manifests itself at key times across a research project that spans hundreds of facilities across the globe, that becomes a massive, worldwide HPC cloud computing challenge.
As such, the ATLAS project was invited by Google to test the Google Compute Engine in an effort to complete that challenge.
The experience has gone well so far, according to Panitkin. “All in all, we had a great experience with Google Computing. We tested several computational scenarios on that platform…we think that Google Compute Engine is a modern cloud infrastructure that can serve as a stable, high performance platform for scientific computing.”
The ATLAS collector, diagrammed below, was designed to intake and record 800 million proton-proton interactions per second. Of those 800 million collisions per second, only about 0.0002 Higgs signatures are detected per second. That translates to one signature for every 83 minutes or so. The computing systems have to sift through that huge dataset containing information from each of those almost billion interactions a second to find that one distinct pattern.
Thankfully, much of the ATLAS data is instantly filtered and discarded by an automatic trigger system. Were this not the case, the collector would generate a slightly unsustainable petabyte of data per second.
Adding to the challenge that the enormous amount of data presents is the very particular signature the ATLAS project was looking for. According to Panitkin, sifting through that much is akin to trying to find just one person in a system of a thousand planets of the same population as Earth. To help visualize what that looks like, the above picture represents all the possible signatures while the diagram below shows the one specific indicator of the Higgs Boson.
CERN collects the data and initially distributes it to its 11 tier-one centers, as shown in the diagram below. The cloud and specifically the Google Compute Engine enter the picture in tier two, where about two hundred centers across the globe simulate their respective sections based on the tier-naught CERN data.
Combining all of those resources into a shared system is essential for scientific researchers, as they cull information from other tests and simulations run. According to Andrew Hanushevsky, who presented alongside Panitkin at the Google I/O event, the system was aggregated using the XRootD system. XRootD, coupled with cmsd, was instrumental and combining and managing the thousand-core PROOF cluster made for ATLAS as well as the 4000-core HTCondor cluster for CERN’s collision analysis.
The important aspect was ensuring the system acted as one, as Hanushevsky explained. “This is a B tree, we can split it up anyway we want and this is great for doing cloud deployment. Part of that tree can be inside the GCE, another part can be in a private cloud, another part in a private cluster, and we can piece that all together to make it look like one big cluster.”
With that in place, the researchers could share information across the network at an impressive transfer rate of 57 Mbps transfer rate to the Google Compute Engine.
Finally, according to Panitkin, the computations done over GCE were impressively accurate. The system reported, according to Panitkin, “no failures due to Google Compute Engine.”
The best science requires extensive collaboration. Global projects such as the one that found the Higgs Boson mark the pinnacle of that collaboration, and these efforts can only grow stronger with the betterment of large-scale cloud-based computing services like Google Compute Engine.