Kris Meier November 17, 2019
We just announced the availability of our Application Accelerator. If you’ve missed it, you can read the news release.
Our Application Accelerator dramatically reduces application latency by accelerating file metadata requests at the speed of DRAM. By independently processing file metadata, InfiniteIO’s Application Accelerator delivers faster application run times resulting in quicker time to market and greater productivity. In metadata-intensive workloads we’ve seen an increase in application performance of 50% or more.
Key technology innovation is the InfiniteIO Metadata Engine (IME). IME processes metadata requests in real-time, responding directly from the network instead of the Network Attached Storage. This abstraction of metadata processing from data processing reduces latency from milliseconds to microseconds.
Why are we focusing on file metadata?
Incremental improvements in storage and interconnect technology clearly enable more file I/Os but they miss the bigger picture, inherent protocol overhead. NAS protocols are chatty, many requests for file metadata must be processed before the NAS storage system can perform reads and writes. All this metadata processing leads to increased application latency. InfiniteIO’s Application Accelerator is the world’s fastest file metadata platform, eliminating unnecessary metadata latency by responding to requests in under 100 microseconds.
Why File Metadata Offload?
Under the covers, applications are generating file metadata operations – there can be up to 7 metadata requests or even more depending on the workload before an actual read or write.
A simple read of a file at /mnt/homedirs/Bob/test/debug.txt results in the generation of metadata requests for directory location then access for each directory en route to the file to be opened. This read generates 7 file metadata operations before the first Read or 87.5% of file operations are really metadata operations.
In metadata intensive workloads such as AI/ML, chip design, genomics, software development, the ratio of metadata operations to data operations can be much higher. A recent exercise in machine learning training picture classification models showed 19 of 20 operations were, in fact, metadata operations. That’s 95% of all file operations are really for metadata! Millions of metadata operations are generated as pictures used in training are discovered, then opened and read. Under the covers the OS is looking up the location of each file and validating client access rights along the way. In NFS, this predominantly generates a series of GETATTR, LOOKUP, and ACCESS calls — all of which InfiniteIO’s Application Accelerator offloads and accelerates at 10x-20x faster than a typical NAS storage system.
In these workloads we’ve seen up to 50% or more increase in application performance by offloading and accelerating file metadata.
In the example above the 19 of 20 represents the number of metadata operations for each read operation in that ML test. Said differently, InfiniteIO was able to offload and accelerate 19 file operations before sending a single operation to the NAS storage system to process. Think of this as a “metadata offload ratio” of 19:1 for this particular ML training test. InfiniteIO is processing 19 file operations for every one operation the NAS storage system is serving. Our Application Accelerator is processing these requests at under 100 microseconds, faster than any storage system on the planet, resulting in a 50% reduction in training time as we are keeping the Application CPUs/GPUs fed with data as fast as possible to train the classification model.
What about other workloads?
We’ve just scratched the surface of what we can do. Early observations and testing show significant metadata offload ratios in the following areas:
- Office productivity software showed a metadata offload ratio of 33:1
- Software build 400:1, and
- Source code control of 1,300:1.
Overall resulting in application acceleration of 50% or more.
Metadata Offload is the next performance frontier.
NAS storage systems historically have addressed performance improvements by making incremental gains in faster media and interconnects to speed up reads and writes. Metadata performance clearly also benefits from these improvements. But not enough. Burdening a filesystem with 1,300 metadata operations for each READ or WRITE increases application latency, slowing everything down. Only InfiniteIO’s Application Accelerator offloads those metadata operations and responds at the speed of DRAM, dramatically decreasing application latency resulting in reduced run times, faster time to market and increased productivity.