Webster Home Page

6.8 NUMA and Peripheral Devices

Although most of the RAM memory in a system is based on high-speed DRAM interfaced directly to the processor's bus, not all memory is connected to the CPU in this manner. Sometimes a large block of RAM is part of a peripheral device and you communicate with that device by writing data to the RAM on the peripheral. Video display cards are probably the most common example, but some network interface cards and USB controllers also work this way (as well as other peripherals). Unfortunately, the access time to the RAM on these peripheral devices is often much slower than access to normal memory. We'll call such access NUMA1 access to indicate that access to such memory isn't uniform (that is, not all memory locations have the same access times). In this section we'll use the video card as an example, although NUMA performance applies to other devices and memory technologies as well.

A typical video card interfaces to the CPU via the AGP or PCI (or much worse, ISA) bus inside the computer system. The PCI bus nominally runs at 33 MHz and is capable of transferring four bytes per bus cycle. In burst mode, a video controller card, therefore, is capable of transferring 132 megabytes per second (though few would ever come close to achieving this for technical reasons). Now compare this with main memory access. Main memory usually connects directly to the CPU's bus and modern CPUs have a 400 MHz 64-bit wide bus. Technically (if memory were fast enough), the CPU's bus could transfer 800 MBytes/sec. between memory and the CPU. This is six times faster than transferring data across the PCI bus. Game programmers long ago discovered that it's much faster to manipulate a copy of the screen data in main memory and only copy that data to the video display memory when a vertical retrace occurs (about 60 times/sec.). This mechanism is much faster than writing directly to the video memory every time you want to make a change.

Unlike caches and the virtual memory subsystem that operate in a transparent fashion, programs that write to NUMA devices must be aware of this and minimize the accesses whenever possible (e.g., by using an off-screen bitmap to hold temporary results). If you're actually storing and retrieving data on a NUMA device, like a Flash memory card, then you must explicitly cache the data yourself. Later in this text you'll learn about hash tables and searching. Those techniques will help you create your own caching system for NUMA devices.

6.9 Segmentation

Segmentation is another memory management scheme, like paging, that provides memory protection and virtual memory capabilities. Linux and Windows do not support the use of segments, nor does HLA provide any instructions that let you manipulate segment registers or use segment override prefixes on an instruction2. These 32-bit operating system employ the flat memory model that, essentially, ignore segments on the 80x86. Furthermore, the remainder of this text also ignores segmentation. What this means is that you don't really need to know anything about segmentation in order to write assembly language programs that run under modern OSes. However, it's unthinkable to write a book on 80x86 assembly language programming that doesn't at least mention segmentation. Hence this section.

The basic idea behind the segmentation model is that memory is managed using a set of segments. Each segment is, essentially, its own address space. A segment consists of two components: a base address that contains the address of some physical memory location and a length value that specifies the length of the segment. A segmented address also consists of two components: a segment selector and an offset into the segment. The segment selector specifies the segment to use (that is, the base address and length values) while the offset component specifies the offset from the base address for the actual memory access. The physical address of the actual memory location is the sum of the offset and the base address values. If the offset exceeds the length of the segment, the system generates a protection violation.

Segmentation on the 80x86 got a (deservedly) bad name back in the days of the 8086, 8088, and 80286 processors. The problem back then is that the offset into the segment was only a 16-bit value, effectively limiting segments to 64 kilobytes in length. By creating multiple segments in memory it was possible to address more than 64K within a single program; however, it was a major pain to do so, especially if a single data object exceeded 64 kilobytes in length. With the advent of the 80386, Intel solved this problem (and others) with their segmentation model. By then, however, the damage had been done; segmentation had developed a really bad name that it still bears to this day.

Segments are an especially powerful memory management system when a program needs to manipulate different variable sized objects and the program cannot determine the size of the objects before run time. For example, suppose you want to manipulate several different files using the memory mapped file scheme. Under Windows or Linux, which don't support segmentation, you have to specify the maximum size of the file before you map it into memory. If you don't do this, then the operating system can't leave sufficient space at the end of the first file in memory before the second file starts. On the other hand, if the operating system supported segmentation, it could easily return segmented pointers to these two memory mapped files, each in their own logical address space. This would allow the files to grow to the size of the maximum offset within a segment (or the maximum file size, whichever is smaller). Likewise, if two programs wanted to share some common data, a segmented system could allow the two programs to put the shared data in a segment. This would allow both programs to reference objects in the shared area using like-valued pointer (offset) values. This makes is easier to pass pointer data (within the shared segment) between the two programs, a very difficult thing to do when using a flat memory model without segmentation as Linux and Windows currently do.

One of the more interesting features of the 80386 and later processors is the fact that Intel combined both segmentation and paging in the same memory management unit. Prior to the 80386 most real-world CPUs used paging or segmentation but not both. The 80386 processor merged both of these memory management mechanisms into the same chip, offering the advantages of both systems on a single chip. Unfortunately, most 32-bit operating systems (e.g., Linux and Windows) fail to take advantage of segmentation so this feature goes wasted on the chip.

6.10 Putting it All Together

CPU architects divide memory into several different types depending on cost, capacity, and speed. They call this the memory hierarchy. Many of the levels in the memory hierarchy are transparent to the programmer. That is, the system automatically moves data between levels in the memory hierarchy without intervention on the programmer's part. However, if you are aware of the effects of the memory hierarchy on program performance, you can write faster programs by organizing your data and code so that it conforms to the expectations of the caching and virtual memory subsystems in the memory hierarchy.

1Remember, NUMA stands for NonUniform Memory Access.

2Though you could easily create macros to do this.

Web Site Hits Since
Jan 1, 2000