x86 Scale Up

I have been introduced to a very cool software/hardware combination yesterday. It has been, without exaggerating, one of the coolest things I have seen in a while.

As you may know, x86 has an issue with scaling up. It’s that x86 architectures and price don’t justify scaling up to tenths and hundreds of CPUs. The multi-core technology introduced in the last few years made a four-way server seem trivial today, where in the past it was a high-performance server for large (and expensive) data centers. It is very common today to purchase an eight-way server at a price of a mere commodity server – all thanks to the multi-core technology.

However, when compared to the large Unix data centers, where 64 and 128 cpus are rather common (I will emphasis – the large Unix data centers), although nowadays, per-core, x86 is somewhat more powerful, for a large load set, it could not rival any many-way server. The common solution with x86 was to “scale out” – add more cheap servers and manage the workload in a more distributed way. Yes, you might pay with communication overhead, however, this can be made cheaper still.

With a distributed load sharing came the illnesses of communication latencies. Myrinet, 10Gb/s Ethernet and Infiniband were a common, yet expensive (as it was a niche market) solutions, and still – for distribution of high loads, they were well worth it. Still – a large scaled-up server based on x86 was nowhere to find.

No more. With ScaleMP’s concatenation you can “bundle” a set of servers using Infiniband link into a single huge-multi-way, huge-ram server at a very low cost, relatively.

Think about how you can purchase your current server, for example, your eight-core server (two quad-core cpus), and in time, scale it up into more powerful server (add another two quad-core cpus), or add more RAM, or more network interfaces, or whatever.

This is not as fast as the IBM x3950 board-link (excuse me for not knowing the exact name), so it is not ideal for databases or systems which tend to create a lot of cache-misses, however for large (actually – very large) SMP systems, it could be great. It can allow any company which feels that the current server might not be enough the safety and assurance that they can actually scale up, using the same server, into adding more cpus and more RAM to the server at any time.

I is supported, as far as I know, only for Linux at the time being. It diminishes some of the distance between the large Unix machines and the modern Linux, for a fracture of the price.

I liked it.