ProfitBricks, a new supplier of infrastructure-as-a-service (IaaS), launched Monday, claiming it can increase the size of servers on demand--scaling up vertically--rather than invoking the usual public infrastructure technique of scaling out horizontally, adding more virtual server nodes.

The difference may seem small, but some applications, particularly large database systems, scale up more easily than they scale out. They function most efficiently when they can command all the resources of a larger server rather than being spread across many small servers. "Most applications and services are meant to scale vertically," said Bob Rizika, CEO of Berlin-headquartered ProfitBricks' U.S. division, in an interview.

The firm announced its IaaS May 21 in Berlin, where it operates a data center. The U.S. unit offers the service out of its data center in Las Vegas. Rizika said this is "second generation infrastructure-as-a-service," different from leading first-generation suppliers and better able to match what customers have in their own data centers.

For example, the maximum server size available as standard infrastructure from Amazon Web Services is an extra-large server with four virtual CPUs amounting to eight Amazon EC2 compute units (ECUs). It also offers a high-memory, quadruple extra-large server with eight virtual CPUs amounting to 26 ECUs. An ECU is the equivalent of a 2007 Intel Xeon or AMD Opteron CPU running at 1 GHz.

[ Want to learn more about how competition continues to drive down pricing in the cloud? See Cloud's Thorniest Question: Does It Pay Off? ]

Rizika said ProfitBricks' minimum compute unit is one current AMD core with a processing power equal to four ECUs. ProfitBricks decided not to subdivide cores into virtual CPUs to avoid the noisy neighbor problem where there's contention for a core's buffers, caches, and I/O resources between a particularly active application and the other virtual machines on a server. For customers wishing to scale up, ProfitBricks' infrastructure may produce servers with up to 48 AMD cores, or the equivalent in Rizika's interpretation of 192 ECUs--all on a single physical host.

Amazon, of course, can provide an equal amount of compute power, but it would be spread across multiple extra-large virtual machines with the customer not knowing how many physical hosts it was on.

To be able to reach the 48-core size, ProfitBricks produced its own virtual machine management software that assigns more cores to a running virtual machine without disrupting its operation. If the host server doesn't have enough free cores to satisfy the need, the workload is transferred to another host, which does.

Rizika and Andreas Gauger, chief marketing officer and co-founder of ProfitBricks, said the server scale-up is handled smoothly enough that users' work on a running application is not disrupted. They add there might be several seconds where a user might suspect his virtual machine had slowed down or was responding sluggishly as the migration occurs. The same claims are made by VMware and Citrix for VMware's ESX Server and Citrix XenServer's live migrations.

In effect, ProfitBricks says it is scaling up a virtual server with a running workload by providing live migration for its version of KVM when it's needed. "We've done it on our own," in the process of designing and building its IaaS over the last two years, said Gauger during the interview. The firm uses "a heavily modified" form of Red Hat's KVM open source hypervisor in its infrastructure.

Because it can scale up in this fashion, ProfitBricks is departing from the pattern of offering a handful of pre-selected server sizes, such as micro, small, large, and extra large. It will allow a user to build a virtual server with the exact number of cores it decides is right for the job, up to the maximum 48. Such a server can be assigned up to 196 GB of RAM. Rackspace, Microsoft, and Amazon Web Services all simplify their infrastructure offerings by offering a limited number of server classes, with predetermined numbers of CPU and amounts of RAM.

To assist users in designing their servers, ProfitBricks is offering a visual user interface where customers set the number of CPUs, amount of storage, and network bandwidth. The Data Center Design user interface lets a customer design a set of servers with volumes of storage and I/O connections on what ProfitBricks calls "a virtual whiteboard."

The design may include firewalls, load balancers, and network connections; the customer may drag and drop lines between servers and the available networks contained on a map of the existing data center infrastructure produced by the interface. When the virtual whiteboard diagram is complete, the customer fires up the systems with a mouse click. The servers, storage, and network connections are automatically loaded into tables to record the design.

ProfitBricks' minimum charge is six cents an hour for a small server with a single virtual CPU (one AMD core). It will compile charges by the minute, instead of by the hour, to keep customers' bills closer to actual amounts of time used.

ProfitBricks' networking is a based on the Infiniband protocol, which offers bigger pipes than 10 Gbps Ethernet, but customers may communicate with it over Ethernet. All communications between enterprise customers and its data center will be over VPNs.

Rizika said the service has been in customer beta, then customer preview for the past year. It became generally available with Monday's announcement.

"The customers moving to the cloud today want to replicate their data center in the cloud ... Our goal is to bring all these features to market and do it at competitive prices," said Rizika. The firm has 200 infrastructure customers, he said.