By Timothy Prickett Morgan
The new cut of IBM Corp’s Unix operating system, AIX Version 4.3.3, is more than just a rehash that has been tweaked to support 24-way symmetric multiprocessing. It is also more than just the code base that will reportedly represent about 80% of the code that is going into the IBM-SCO Monterey/64 Unix variant for Intel’s forthcoming IA-64 chips. AIX 4.3.3 is the version of AIX that IBM should have created years and years ago and has finally built.
AIX 4.3.3, like prior V4 releases, runs on 32-bit PowerPC and 64- bit PowerPC and Power3 processors in the current RS/6000 line of workstations and servers. It will also run on the original Power processors announced in 1990 as well as on the subsequent Power2 processors that first appeared in the mid-1990s. AIX 4.3.3 supports up to 64Gb of main memory, double that of AIX 4.3.2, and supports 24-way SMP compared to the 12-way SMP in AIX 4.3.2. While these are important, perhaps the most significant addition to AIX with V4.3.3 is a feature called Workload Management. While Workload Management falls short of the hardware and logical partitioning found in IBM AS/400s and mainframes, it is a significant improvement over past AIXes, which essentially had no features to help system administrators keep applications from butting heads and running away with system resources.
IBM AS/400s and mainframes were designed from the ground up to be multipurpose machines forsupporting simultaneous, mixed workloads on a single machine, Unix servers from IBM and other vendors have been relatively poor at handling mixed workloads. The reason why server farms proliferate in the Windows NT space is that NT is essentially incapable of handling more than one enterprise-class application at a time; you basically need multiple servers for each application workload you want to support. And they multiply fast. You need a few for scalability because NT servers haven’t scaled well, and you need to double up again for system mirroring because you can’t trust NT servers (mainly because of the software, not the hardware) to stay running for more than a few weeks without crashing mightily. While the Unix servers from IBM, Hewlett-Packard Co, Sun Microsystems Inc, Compaq Computer Corp and others are certainly much better than NT in the reliability department, they still can’t do much more than one, two or three things at a time. In most cases, big Unix shops buy a rack of servers and spread their applications across them because it is just easier than trying to explain to end users and the CEO why some rogue SQL query has brought the order entry system to its knees at the height of the Christmas buying season.
AIX Workload Management gives AIX the kind of software subsystem capability that has been the hallmark of the AS/400 minicomputer, although it is unclear if the RS/6000 division actually solicited advice and/or code from the AS/400 division while it was creating the Workload Management feature. Administrators set up Workload Management by defining specifically what processor and memory resources can be allocated to particular jobs or classes of jobs. The software allows administrators to set both resource minimums and maximums for up to 29 different classes of jobs in nine different priority tiers on a single machine. This is not the same thing as the dynamic domains in Sun’s Enterprise 10000 servers, but it is very similar to the fine-grain control that Sun has in Solaris Resource Manager. AIX Workload Management falls short of the logical partitioning that IBM has built into AS/400s with OS/400 V4R4, announced earlier this year with the Northstar AS/400e servers. With V4R4, customers with a symmetric multiprocessor can carve up processors, memory, and I/O in their machines into virtual individual AS/400s that run their own copies of the OS/400 operating system and can run in different time zones with different languages.
On the IBM mainframe line, logical partitioning goes even further, allowing customers to put up to 15 logical partitions on a single machine with 12 real processors. The mainframe partitions can be weighted according to whatever processing resources customers want to allocate to the partitions they carve up; on the AS/400, you allocate processors and that’s all the fine tuning you get. The mainframe allows for dynamic allocation of resources across logical partitions – this feature will be added to future releases of OS/400, perhaps V4R5 in the spring. Sun Dynamic Resource Management can only reallocate resources within a physical hardware partition on a Starfire server. HP will in the first half of next year add logical partitioning to HP-UX 11, and by the end of the year hopes to have features that allow its HP 9000 server customers to manage the resources in multiple domains; how dynamic the reallocation of resources will be within HP-UX 11 remains to be seen. IBM’s approach with AIX 4.3.3 attacks the dynamic reallocation problem directly rather than indirectly by letting applications only have so much memory and so many processor cycles out of the total in the machine. How I/O gets allocated is unclear, and so is the impact that the omission of I/O allocation has on the effectiveness of the Workload Management feature. The goals of the AS/400 logical partitioning design were to promote data center consolidation for multinational customers, server consolidation for various jobs within a single location and giving developers a slice of the machine so they could test programs on the machine they will eventually run on. Since IBM already has a partitionable RS/6000 server – the SP parallel server – it did not have to go that far with AIX 4.3.3. Workload Management functions are controlled through what IBM calls a Web-based System Manager (WebSM), SMIT, shell scripts or command line interfaces.
IBM also made lots of smaller tweaks to AIX with Version 4.3.3, including support for Sun’s NIS+ system management programs, a journal file system that now supports online backup as well as concurrent data mirroring and data striping on file systems, upgrded Sendmail, and tweaks to the kernel to improve web serving performance. The new software also includes Cisco’s EtherChannel network bandwidth aggregation and load balancing software that augments the Fast Ethernet 802.3 standard. AIX 4.3.3 also includes a whole slew of TCP/IP and other communication enhancements, and an improved AIX Fast Connect, which offers print file serving for Windows and OS/2 clients. Finally, the AIX graphics subsystems have been enhanced with the X11R6.3 Broadway release from the X Consortium. Geographic Remote Mirroring, or GeoRM, has been included as an add-on feature to AIX 4.3.3.
