Recent from talks
Virtual memory compression
Knowledge base stats:
Talk channels stats:
Members stats:
Virtual memory compression
Virtual memory compression (also referred to as RAM compression and memory compression) is a memory management technique that utilizes data compression to reduce the size or number of paging requests to and from the auxiliary storage. In a virtual memory compression system, pages to be paged out of virtual memory are compressed and stored in physical memory, which is usually random-access memory (RAM), or sent as compressed to auxiliary storage such as a hard disk drive (HDD) or solid-state drive (SSD). In both cases the virtual memory range, whose contents has been compressed, is marked inaccessible so that attempts to access compressed pages can trigger page faults and reversal of the process (retrieval from auxiliary storage and decompression). The footprint of the data being paged is reduced by the compression process; in the first instance, the freed RAM is returned to the available physical memory pool, while the compressed portion is kept in RAM. In the second instance, the compressed data is sent to auxiliary storage but the resulting I/O operation is smaller and therefore takes less time.
In some implementations, including zswap, zram and Helix Software Company’s Hurricane, the entire process is implemented in software. In other systems, such as IBM's MXT, the compression process occurs in a dedicated processor that handles transfers between a local cache and RAM.
Virtual memory compression is distinct from garbage collection (GC) systems, which remove unused memory blocks and in some cases consolidate used memory regions, reducing fragmentation and improving efficiency. Virtual memory compression is also distinct from context switching systems, such as Connectix's RAM Doubler (though it also did online compression) and Apple OS 7.1, in which inactive processes are suspended and then compressed as a whole.
There are two general types of virtual memory compression : (1) sending compressed pages to a swap file in main memory, possibly with a backing store in auxiliary storage, and (2) storing compressed pages side-by-side with uncompressed pages.
The first type (1) usually uses some sort of LZ class dictionary compression algorithm combined with entropy coding, such as LZO or LZ4, to compress the pages being swapped out. Once compressed, they are either stored in a swap file in main memory, or written to auxiliary storage, such as a hard disk. A two stage process can be used instead wherein there exists both a backing store in auxiliary storage and a swap file in main memory and pages that are evicted from the in-memory swap file are written to the backing store with a much increased write bandwidth (eg. pages/sec) so that writing to the backing store takes less time. This last scheme leverages the benefits of both previous methods : fast in-memory data access with a large increase in the total amount of data that can be swapped out and an increased bandwidth in writing pages (pages/sec) to auxiliary storage.
One example of a class of algorithms for type (2) virtual memory compression is the WK (Wilson-Kaplan et. al) class of compression algorithms. These take advantage of in-memory data regularities present in pointers and integers. Specifically, in (the data segment -- the WK algorithms are not suitable for instruction compression) target code generated by most high-level programming languages, both integers and pointers are often present in records whose elements are word-aligned. Furthermore, the values stored in integers are usually small. Also pointers close to each other in memory tend to point to locations that are themselves nearby in memory. Additionally, common data patterns such as a word of all zeroes can be encoded in the compressed output by a very small code (two bits in the case of WKdm). Using these data regularities, the WK class of algorithms use a very small dictionary ( 16 entries in the case of WKdm ) to achieve up to a 2:1 compression ratio while achieving much greater speeds and having less overhead than LZ class dictionary compression schemes.
By reducing the I/O activity caused by paging requests, virtual memory compression can produce overall performance improvements. The degree of performance improvement depends on a variety of factors, including the availability of any compression co-processors, spare bandwidth on the CPU, speed of the I/O channel, speed of the physical memory, and the compressibility of the physical memory contents.
On multi-core, multithreaded CPUs, some benchmarks show performance improvements of over 50%.
Hub AI
Virtual memory compression AI simulator
(@Virtual memory compression_simulator)
Virtual memory compression
Virtual memory compression (also referred to as RAM compression and memory compression) is a memory management technique that utilizes data compression to reduce the size or number of paging requests to and from the auxiliary storage. In a virtual memory compression system, pages to be paged out of virtual memory are compressed and stored in physical memory, which is usually random-access memory (RAM), or sent as compressed to auxiliary storage such as a hard disk drive (HDD) or solid-state drive (SSD). In both cases the virtual memory range, whose contents has been compressed, is marked inaccessible so that attempts to access compressed pages can trigger page faults and reversal of the process (retrieval from auxiliary storage and decompression). The footprint of the data being paged is reduced by the compression process; in the first instance, the freed RAM is returned to the available physical memory pool, while the compressed portion is kept in RAM. In the second instance, the compressed data is sent to auxiliary storage but the resulting I/O operation is smaller and therefore takes less time.
In some implementations, including zswap, zram and Helix Software Company’s Hurricane, the entire process is implemented in software. In other systems, such as IBM's MXT, the compression process occurs in a dedicated processor that handles transfers between a local cache and RAM.
Virtual memory compression is distinct from garbage collection (GC) systems, which remove unused memory blocks and in some cases consolidate used memory regions, reducing fragmentation and improving efficiency. Virtual memory compression is also distinct from context switching systems, such as Connectix's RAM Doubler (though it also did online compression) and Apple OS 7.1, in which inactive processes are suspended and then compressed as a whole.
There are two general types of virtual memory compression : (1) sending compressed pages to a swap file in main memory, possibly with a backing store in auxiliary storage, and (2) storing compressed pages side-by-side with uncompressed pages.
The first type (1) usually uses some sort of LZ class dictionary compression algorithm combined with entropy coding, such as LZO or LZ4, to compress the pages being swapped out. Once compressed, they are either stored in a swap file in main memory, or written to auxiliary storage, such as a hard disk. A two stage process can be used instead wherein there exists both a backing store in auxiliary storage and a swap file in main memory and pages that are evicted from the in-memory swap file are written to the backing store with a much increased write bandwidth (eg. pages/sec) so that writing to the backing store takes less time. This last scheme leverages the benefits of both previous methods : fast in-memory data access with a large increase in the total amount of data that can be swapped out and an increased bandwidth in writing pages (pages/sec) to auxiliary storage.
One example of a class of algorithms for type (2) virtual memory compression is the WK (Wilson-Kaplan et. al) class of compression algorithms. These take advantage of in-memory data regularities present in pointers and integers. Specifically, in (the data segment -- the WK algorithms are not suitable for instruction compression) target code generated by most high-level programming languages, both integers and pointers are often present in records whose elements are word-aligned. Furthermore, the values stored in integers are usually small. Also pointers close to each other in memory tend to point to locations that are themselves nearby in memory. Additionally, common data patterns such as a word of all zeroes can be encoded in the compressed output by a very small code (two bits in the case of WKdm). Using these data regularities, the WK class of algorithms use a very small dictionary ( 16 entries in the case of WKdm ) to achieve up to a 2:1 compression ratio while achieving much greater speeds and having less overhead than LZ class dictionary compression schemes.
By reducing the I/O activity caused by paging requests, virtual memory compression can produce overall performance improvements. The degree of performance improvement depends on a variety of factors, including the availability of any compression co-processors, spare bandwidth on the CPU, speed of the I/O channel, speed of the physical memory, and the compressibility of the physical memory contents.
On multi-core, multithreaded CPUs, some benchmarks show performance improvements of over 50%.