Making the Edge Efficient with Unum!

Unum, Gcore, and Intel partner to bring the most efficient data-processing infrastructure to the edge

In 2023, software scaling is often synonymous with the cloud, wherein a vast majority of commodity servers are connected over a relatively poor fabric.

This misconception related to software scaling introduces several performance issues.

1. Most companies store their data in one central location, resulting in up to 300 ms access latencies for some regions.
2. Within that data center, the information is often sharded into thousands of servers with virtualized storage capable of only ~50,000 operations/second/node.

Every time you update your enterprise data, the signal would have to travel halfway around the world, and would end up synchronizing countless tiny virtual machines deployed on weak and outdated hardware.

Most cloud customers are forced to buy products with advertised infinite horizontal scalability, resulting in additional costs and latency penalties. But in reality, modern SSDs can reach 1.5 million operations/second, and the customers often wouldn’t even need that scaling if the software was well optimized.

As a result, a partnership has been formed among Unum, Gcore and Intel, with a collective goal to achieve the fastest edge storage systems, targeting a minimum of 10 million operations per second. This will be made possible by utilizing Intel’s latest hardware, leveraging Gcore’s infrastructure, and incorporating the Unum transactional database technology. By doing so, we aim to provide enterprises and telecommunications companies with unparalleled access to the exceptionally high-speed edge technology.

Optimizing the nodes

Most modern IO is built on layers of costly abstractions. Users of Postgres and other conventional databases have to experience multiple levels of indirections, mutexes, copies, and other bottlenecks while passing data to and from storage devices.

Several lighter alternatives have been developed over the last 5 years, aligned with the development of UStore, Unum’s Multi-Modal Transactional Database. The most important ones are io_uring and SPDK. With those, UStore can avoid expensive system calls and partially or entirely bypass the OS and Filesystem layers. To achieve that, all components of a DBMS have to be replaced. Most importantly, the underlying persistent datastructure – the Key-Value Store (KVS).

Most modern DBMS brands, like CockroachDB, Yugabyte, and Neo4J rely on Facebook’s Open-Source KVS called RocksDB. Unum uses an in-house technology called UDisk. Both implement Log Structured Merge (LSM) Trees, the write-optimized alternative to conventional persistent B-Trees. LSMs avoid small random writes, by packaging all updates together in-memory and in the Write-Ahead-Log, and submitting in batches. Fast as it is, the procedure produces redundant copies of data, which have to be compacted over time. It’s the database analog of Garbage Collection. With UDisk we can exploit hardware-acceleration for such operations, leveraging both SIMD-assembly for x86 AVX2 and AVX-512, as well as Arm Neon and SVE. UDisk is also the first transactional system to offload those “compactions” onto a GPU in the system’s presence of an underutilized Nvidia card.

UDisk has so far outperformed RocksDB on every CPU and every NVME SSD we have seen, across on-Prem, Cloud, and Edge installations, making it ideal for an efficient low-latency DBMS product – UStore. UStore complements the speeds of UDisk with the functionality rivaling Redis, Neo4J, and MongoDB at the same time. It supports JSON, BSON, MessagePack and other flexible-schema Document Collections, Graph Collections, Binary Collections, Apache Arrow interoperability, and an entire ecosystem of tools and bindings for C++, Python, Java, GoLang.

UStore was designed for the age of AI with batch workloads in mind. It can uniquely connect with PyTorch and other Machine Learning pipelines, performing random-sampling directly from your main OLTP DBMS, vastly simplifying the data-exploration. It also natively supports Vector Search, and brings an entire family of bundled Transformer models optimized for cheap and efficient inference on Edge, forming the first complete Database+AI product lineup for customers worldwide!

Optimizing the hardware

Economic recessions always force the companies to be more efficient, cut costs, and replace outdated overpriced enterprise software. That procedure, however, is different for startups and centi-billion dollar multinational corporations. A single Telecom can spend over $1B a year on electricity, even with today’s minimal 5G adoption. On that scale, working with hardware partners is crucial to maximize our customers’ savings and even global CO2 emissions.

We are excited to partner with Intel around their Sapphire Rapids CPU platform and Optane SSDs. With Sapphire Rapids we not only gain access to DDR5, and PCIE Gen 5, but also broader AVX-512 functionality and AMX instructions for faster matrix multiplications. With Optane we get a storage medium so fast that we can halve the number of servers and their internal RAM capacity, while still sustaining much higher throughput than Edge location today!

Optimizing the deployments

In the world of databases, very few can fully utilize the immense throughput provided by cutting-edge hardware. However, Unum UStore stands apart from the rest.

Similarly, when it comes to bringing high-end equipment to the edge, most cloud providers fall short. But Gcore is different. Gcore boasts the powerful SPR server, built on Quanta hardware—Intel Xeon Platinum 8480+. With this exceptional hardware, Unum UStore and Gcore are revolutionizing the possibilities at the edge.

Why did Unum choose Gcore?

When selecting a partner for the project, Gcore emerged as the clear choice due to their impressive track record and expertise in various technological solutions. As a renowned global leader in public cloud and edge computing, content delivery, hosting, and security, Gcore has established itself as a trusted name in the industry.

One of the key factors that made Gcore stand out is the exceptional network capacity and connectivity. With the rise of 5G technology, Gcore has been at the forefront, providing seamless 5G connectivity to cloud infrastructure. This capability ensures that our project benefits from the latest advancements in network technology, enabling efficient and lightning-fast data transfer. Furthermore, Gcore’s advanced DDoS protection adds an extra layer of security to our project.

Beyond technical capabilities, we found that Gcore and Unum share a common vision and approach. Our partnership is driven by a shared mindset and chemistry, making collaboration seamless and productive. Gcore’s commitment to innovation and excellence aligns perfectly with Unum’s goals, allowing us to work together seamlessly towards achieving our objectives.

Notably, Gcore’s global presence and extensive network infrastructure play a crucial role in our decision. With over 140 points of presence in reliable Tier IV and Tier III data centers worldwide, Gcore effectively manages its own IT infrastructure across six continents. This ensures reliable and robust service delivery, reinforcing our confidence in their ability to support our project’s requirements.

We always aim to unite with the world’s best hardware vendors to bring the most performant managed database service to 140 countries! For you, the end customer, that would mean:

• Under 30 ms latency to access your nearest location.
• Over 10 M random read operations per node.

Excited? Subscribe to Unum’s and Gcore’s newsletters to learn when these nodes reach your country! Click the subscribe button located at the end of the page.