Computer Science Brochure

As a program that controls and mediates the access of other programs to hardware resources, an Operating System (OS) faces the challenges of reconciling ever-changing program requirements and rapidly advancing hardware technologies. At IBM Research, these challenges are being addressed with a variety of approaches.

K42

The K42 group is developing a new operating system for 64-bit cache-coherent multiprocessors, ranging from small-scale to very large-scale multiprocessors. Each virtual resource (e. g., file, process, memory region) is implemented by independent objects to achieve a high degree of locality, avoid global locks and data structures, and allow the implementation of a resource to be customized to the demands. Applications with special needs, like databases and Web servers, can pick the building blocks (i. e., objects) that implement the resources they use to match their expected demands. K42 currently runs on 64-bit PowerPC and MIPS® systems.

SawMill

The SawMill project addresses the complexity of developing and maintaining a variety of custom operating systems. With the emergence of embedded and personal systems, the need to create operating systems customized to device and application requirements has increased significantly. We first decompose an existing operating system into flexible, reusable components. We then define an architecture upon which efficient and robust operating systems can be composed. This framework is being applied to Linux® to create SawMill Linux which consists of Linux-based components to provide typical system services and general components such as memory, task, device, and access control managers.

Leeds/Linux

The focus of our Leeds/Linux effort is to explore operating system support for secure embedded devices. Our specific target is the IBM 4758 secure coprocessor, a FIPS-140 level 4 device with hardware cryptographic support and physical tamper protection. The project consists of two parts. The first is to make Linux suitable as a stand-alone operating system for small embedded devices. The second deals with the security aspects of devices, including the handling of tamper-responsive features, encryption of file contents, and handling of trust.

AFPA

The AFPA project develops OS extension technology and kernel components for high-performance network servers. The goal is to achieve better server performance by extending the kernel and TCP/IP stack with a cache and application logic. AFPA targets commercial operating systems (Linux, AIX®, and Windows® 2000, S/390) as well as a multiserver OS based on Linux. Performance is achieved by exposing TCP/IP events and a kernel-managed cache directly to server logic expressed as a state machine.

Data Sharing Facility

The target of the Data Sharing Facility (DSF) research is a serverless file system that distributes management over cooperating machines connected by a fast network. DSF is aimed at building a system that can scale from a few to several hundred machines using commodity components. DSF contains four types of entities: clients, file managers, storage managers, and logical volumes. These entities provide a global memory cache, a distributed file management, and a distributed storage repository. Client machine memories are treated as one global cooperative cache, thereby reducing the cost of cache misses.

iSCSI --- SCSI-over-TCP

iSCSI refers to the ability to access SCSI I/O devices that are connected to a host via a network running TCP. As Local Area Network (LAN) speeds reach the gigabit range, the case for connecting block storage over IP networks becomes increasingly attractive. TCP/IP allows the use of standard communication equipment and existing infrastructure management software tools to handle any I/O configuration with low-cost solutions. Prototypes are currently being implemented with two configurations. In one configuration, a client host is connected via gigabit Ethernet running TCP/IP to a target disk controller. In another configuration, there are two Fiber Channel Storage Area Networks (SANs) each connected to a Linux box running SCSI/TCP software.

Flexible Unix Servers

The flexible Unix server group addresses high end enterprise operating systems with focus on AIX and Linux. We are developing a dynamically partionable large scale shared memory system, e.g. SMPs or NUMA systems. Each partition in such a system runs a fault contained instance of a potentially different OS (e.g. AIX, Linux, ..). A partitionable system can address scalability, workload isolation and capacity on demand requirements. We are adapting AIX and Linux on PowerPC and Intel platforms to enable dynamic partitioning through online addition and removal of system resources such as CPUs, memory and I/O without a need to reboot. We further are investigating the automatic resource movement based on workload characteristics, service level agreements and performance monitoring statistics. The prototype hardware allows the sharing of memory which we are exploring for the usage of virtual LAN. We are also working on Linux enterprise readiness by improving kernel scalability and security.

Internet QoS

The Web and Internet together constitute a critical information, entertainment, and commerce infrastructure that is rapidly evolving from a best-effort service model to one in which service differentiation can be provided for users, services, and applications. This service differentiation (also referred to as Quality of Service (QoS)) is in the form of preferential treatment (using priorities, resource reservation, etc.) of one type of user/application traffic over another at the servers, proxies, and network elements that comprise the end-to-end infrastructure. Many ISPs, network carriers, and Web sites are enabling some form of service differentiation and this emphasis is likely to become even stronger as the Internet continues its exponential growth. We are pursuing a number of research directions investigating support for service differentiation on Internet servers and proxies, including flexible and efficient network bandwidth management, dynamic inbound connection rate control for server overload protection, resource management (CPU, network bandwidth) for co-located Web sites and Internet applications on the same server platform, and policy-enabled management for server farms.