CN107784121A - Lowercase optimization method of log file system based on nonvolatile memory - Google Patents

Abstract

The invention relates to a performance optimization method of a log file system based on non-volatile memory. This method connects NVM and DRAM to the CPU memory bus, uses in-place writing for small data updates to minimize memory usage, uses small data logs to perform fine-grained backup of the data to be updated, and records the latest files to be changed Data information; through a fixed-length circular buffer, store the global hash table, record the starting address of each update, and use it for fast indexing of this data page in the next in-place update. Record the starting address of the updated data page for the next quick index of this data page; after the data writing is completed, update the pointer to the data block by updating the metadata, and update the log tail to submit this write as this A marker that the operation is complete. The invention can balance data consistency overhead and performance, reduce space allocation cost, avoid write amplification in the small data update process, and provide strong data consistency guarantee at the same time.

Description

A lowercase optimization method for journaling file system based on non-volatile memory

technical field

The invention relates to a performance optimization method of a log file system based on a non-volatile memory, in particular to an optimization method for writing performance in a small data update scenario.

Background technique

Several non-volatile memory technologies (non-volatile memory, NVM) are currently being researched and developed, including phase change memory (PCM), ferroelectric memory (FeRAM), memristors (Memristors), spin torque transfer memory ( STT-RAM) and 3D XPoint, etc. These NVMs have the advantages of large capacity, high integration, low leakage power consumption, fast access speed, non-volatility, and byte addressability. Since NVM has both the byte-addressable characteristics of DRAM and the persistence of disk, data can be persisted at the memory level in NVM. Persistent memory initially simulates block devices in the form of RAMDISK to be compatible with traditional file systems, so that it can benefit from memory-level persistence. However, the traditional file system cannot be fully applied to NVMM for two main reasons: First, due to the block addressing characteristics of the disk, the traditional file system generally organizes and manages data by block, and cannot utilize NVM bytes when applied to NVM. Addressing to improve system read and write performance. Secondly, most of the optimizations done by the traditional file system are based on the sequential read and write of the disk. However, NVM does not have a big performance difference between random and sequential read and write. Therefore, if a traditional file system is used for NVM, its data management method will bring redundancy and burden to system performance. Therefore, the current file systems bypass the page cache and directly connect the NVM to the processor through the memory bus, and the processor accesses the NVM at byte granularity to build a persistent memory file system.

For a persistent memory file system, the CPU bypasses the DRAM page cache to directly access the NVM, and the data stored in the NVM must be consistent to be valid and accessible. The CPU and its cache hierarchy reorder operations to improve performance. To achieve this, file system operations must be atomic to prevent the system from crashing and leaving data structures in an inconsistent state before all operations are complete. Currently mainstream file systems use some mechanisms to overcome crash consistency issues, such as journaling, copy-on-write (COW) and log structures. However, these methods have drawbacks. Journaling will perform each write operation twice to prevent data loss. When COW updates data, it first copies the original data block and then makes changes. This will cause write amplification for small updates. The log structure is to update After caching first, write them to the target location uniformly, which requires a large block of continuous space during caching.

In addition, by paying attention to the changes in file access modes in different scenarios, it can be found that the read and write operations of small files will have an important impact on the performance improvement of the file system. Due to block addressing and sequential reading and writing, traditional block devices are always limited in terms of performance improvement for small file reading and writing and small updates. With the emergence of NVM, we can use its byte addressing and other properties to reconstruct the data layout and optimize the access strategy, so as to realize a high-performance file system for NVM, especially for the application scenario of updating small data sets.

Contents of the invention

In order to solve the above-mentioned technical problems, the object of the present invention is to provide a lowercase optimization method of a log file system based on non-volatile memory. This method connects NVM and DRAM to the CPU memory bus, uses in-place writing for small data updates to minimize memory usage, uses small data logs to perform fine-grained backup of the data to be updated, and records the latest changes to the file Data information; through a fixed-length circular buffer, store the global hash table, record the starting address of each update, and use it for fast indexing of this data page in the next in-place update. Record the starting address of the updated data page for the next quick index of this data page; after the data writing is completed, update the pointer to the data block by updating the metadata, and update the log tail to submit this write as this A marker that the operation is complete.

The technical solution adopted by the present invention to solve its technical problems is:

A lowercase optimization method for a log file system based on non-volatile memory, including:

In the first step, the system parallels NVM and DRAM and directly connects them to the CPU memory bus. The CPU uses load and store to access NVM, so that NVM exists at the memory level instead of the block device level, fundamentally realizing the performance of persistent memory promote;

In the second step, the system stores data and logs in NVM in a linked list structure, and each inode uses a separate log to support concurrent access. Store the data index in DRAM in the form of radix tree to support fast search. Data is allocated based on pages, and the default size of a data page is 4096 bytes (4KB);

In the third step, when the data is updated, calculate the number of data pages involved in the update. When the update is larger than the size of a data page in the file system, skip directly to the seventh step;

In the fourth step, when the update is smaller than the size of a data page, first add metadata records to the log. In DRAM, we store a global hash table that can be written cyclically. When writing a single data page for the first time, allocate a new page, and record the block number and starting address of the new data page in the hash table. When writing this data page next time, the address can be quickly indexed through the hash table;

The fifth step is to make a fine-grained backup of the data to be updated. If it is the first write, that is, the address of the data page cannot be found in the hash table through the block number, copy the original data to the new page first, and then write Update data; if the address of the data page can be directly indexed in the hash table, then record the latest data information to be changed in the small data log backup. Then write and update the data page in situ;

Step 6: After the update is complete, update the metadata pointer to the data block, update the log tail to submit this write, indicating that the update is complete, and finally update the pointer to the metadata index;

The seventh step, for writing and updating multiple data pages, first write the changed part of the data in the newly allocated space, and then copy the unchanged original data to the new page. After the data is written, we update the metadata and Point the corresponding log entry to the new page, update the log tail to submit, indicating the completion of the update, and finally update the pointer of the metadata index.

Using the present invention can achieve the following beneficial effects: firstly, it fully utilizes the byte addressing characteristics different from NVM and traditional storage devices, and constructs a hybrid memory with DRAM and NVM. Secondly, the present invention can balance data consistency overhead and performance, reduce space allocation costs, avoid write amplification during small data updates, improve system write performance, and provide strong data consistency guarantees after system crashes.

Description of drawings

Figure 1 is a lowercase flowchart of a log file system based on non-volatile memory;

Fig. 2 is a schematic diagram of a small data update structure.

Detailed ways

With reference to Fig. 1, the present invention provides a kind of lowercase optimization based on the log file system of non-volatile memory, comprising:

In the first step, the system parallels NVM and DRAM and directly connects it to the CPU memory bus. The CPU uses load and store to access NVM. Use the copy_from_user_nocache() interface to bypass the cache to ensure that the data to be written is persisted before the relevant metadata is modified;

In the second step, the system stores data and metadata in NVM in a linked list structure, and each inode uses a separate log to support concurrent access. The metadata index is stored in DRAM in the form of radix tree to support fast search. Data is allocated based on pages, and the default size of a data page is 4096 bytes (4KB);

In the third step, when there is an update operation on the data, first calculate the number of data pages involved in this update. When the update is larger than a data page size of the file system, skip directly to step 7;

In the fourth step, when the update is smaller than the size of a data page, the metadata record is first added to the log. In DRAM, there is a global hash table that can be written cyclically. When writing a single data page for the first time, allocate a new page, and record the block number and starting address of the data page in the hash table. When writing this data page next time, the address can be quickly indexed through the hash table;

The fifth step is to make a fine-grained backup of the data to be updated. If it is the first write, that is, the address of the data page cannot be found in the hash table through the block number, copy the original data to the new page first, and then write Update data; if the address of the data page can be directly indexed in the hash table, then record the latest data information to be changed in the small data log backup. Then write and update the data page in situ;

Step 6: After the update is complete, update the metadata pointer to the data block, update the log tail to submit this write, indicating that the update is complete, and finally update the pointer to the metadata index. Figure 2 is a schematic diagram of data page 0 being updated, <0,1> means starting from data page 0, and the number of pages involved in the updated data is 1;

The seventh step is to write and update multiple data pages, first write the changed data in the newly allocated space, and then copy the unchanged original data to the new page. After the data is written, we update the metadata and point the corresponding log entry to the new page, update the log tail to submit, indicating the completion of the update, and finally update the pointer of the metadata index.

Further as a preferred embodiment, the second step is specifically:

For metadata, the selected log-structured file system uses logs to record every operation of metadata. Every update of metadata will add a log entry. After the update, we submit by changing the log tail to the latest entry. Update, our default processor supports 64-bit atomic write, so the atomicity of the update can be guaranteed. The data page size may be 4KB, 2MB and 1GB according to different processors, here we default to 4KB;

Further as a preferred embodiment, before step 4:

When it is judged that the update is within a data page, the metadata record is first added to the log, and the global hash table that can be written cyclically in DRAM can store the address of the latest updated data page in limited space. When the data page is updated for the first time, a new page is allocated, and the block number and starting address of the data page are recorded in the hash table. When this page is written next time, its starting address can be quickly indexed through the hash table;

Further as a preferred embodiment, in the described step five and six processes:

Before writing data, each inode will have the latest updated data records, including the offset and length records of the changed data. For the data pages that can be found in the hash table, allocate a fixed small data log to perform fine-grained backup of each updated data record, use clflush and mfence to flush the data from the CPU cache for persistence, and then restore the Data pages are written and updated in-place. Compared with multiple data page allocations and repeated writing of unmodified data in the traditional copy-on-write method, using this method can improve the write performance of small data. After a system crash or power failure, read the data in the small data log for redo, thereby completing uncommitted operations and preventing data from being in an inconsistent state.

Further as a preferred embodiment, in the step seven, it is specifically:

When it is judged that the data update involves multiple pages, write the changed part of the data in the newly allocated space, then copy the unmodified original data to the new page, update the metadata and log entry pointer to the new page, and update the log tail. Commit for this update. This can ensure that once the system crashes during the update, as long as the metadata log tail is not submitted, we can still find the original data page based on the old metadata, thus ensuring that the data is in a consistent state.

The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the described embodiments, and those skilled in the art can also make various equivalent deformations or replacements without violating the spirit of the present invention. These equivalent modifications or replacements are all within the scope defined by the claims of the present application.

Claims (5)

1. a kind of small letter optimization method of the JFS based on nonvolatile memory, it is characterised in that connect NVM and DRAM It is connected on CPU rambus, small data is updated and write using original place so as to minimize EMS memory occupation, using small data daily record to wanting The data of renewal carry out fine-grained backup, record the newest data message to be changed of file；It is slow by a circulation Area is rushed, deposits global Hash table, records the initial address of each renewal, to this data page when being updated for original place next time Quick indexing；The initial address for updating the data page is recorded, for next time to the quick indexing of this data page；Data write After the completion of, the pointer of data block is pointed to by more new metadata, Update log tail is to specifically writing into capable submission, as this behaviour Make the mark completed；Comprise the following steps：

The first step, system is arranged side by side with DRAM by NVM, is directly connected on CPU rambus, CPU is come to visit using load and store Ask NVM so that NVM is present in internal memory level rather than block device level, realizes the performance boost of internal memory；

Data and metadata are deposited in NVM by second step, system with list structure respectively, and each inode uses single daily record To support concurrently to access, index of metadata is deposited in into DRAM in the form of radix tree and supports quick lookup, data are based on page Distribution, the size of one data page of acquiescence is 4096 bytes；

3rd step, when data have renewal to operate, calculate this time update involved data number of pages first, when renewal is more than text During one data page size of part system, the 7th step is leapt to；

4th step, when renewal is less than a data page size, metadata record is first added to log, in DRAM, there is one The global Hash table of recyclable write-in, when writing individual data page first, distribute a new page, the record data in Hash table The block number and initial address of page, when writing this data page again next time by Hash table quick indexing to address；

5th step, fine-grained backup first is carried out to the data to be updated, i.e., can not be in Hash by block number if writing first Data page address is found in table, then initial data is first copied into new page, then write and update the data；If direct index in Hash table To data page address, then the newest data message to be changed of file be recorded into small data Log backup, then to data Write renewal in page progress original place；

6th step, after renewal terminates, more new metadata points to the pointer of data block, and Update log tail carries to current row of writing into Hand over, represent that this time renewal is completed, the pointer of final updating index of metadata；

7th step, the renewal of writing for multiple data pages, the data of change part are first write in newly assigned space, then not more The former data duplication changed is to the new page, and after data write-in terminates, our more new metadatas simultaneously point to corresponding daily record entry New page, Update log tail are submitted, and represent the completion of this time renewal, the pointer of final updating index of metadata.

2. a kind of small letter optimization method of JFS based on nonvolatile memory according to claim 1, it is special Sign is that, in described second step, it is specially：

For metadata, the log-structured file system usage log of selection records the operation each time of metadata, Each renewal of metadata can all add a daily record entry, and we are by changing daily record tail to newest after the updating Entry updates to submit.

3. a kind of small letter optimization method of JFS based on nonvolatile memory according to claim 1, it is special Sign is, the 4th described step：

When judging that renewal is within a data page, metadata record is first added to log, in DRAM, has one can follow The global Hash table of ring write-in, the data page address of recent renewal is stored using the confined space, when this data page is to update first When, a new page is distributed, the block number and initial address of record data page, pass through when writing this page again next time in Hash table Hash table quick indexing is to its initial address.

4. a kind of small letter optimization method of JFS based on nonvolatile memory according to claim 1, it is special Sign is, in the 5th described step, six-step process, in addition to：

Before data are write, each inode can have the data record of latest update, including inclined in the page of change data Shifting amount and length records, for the data page found in Hash table, one fixed small data daily record of distribution is to each renewal Data record carries out fine-grained backup, and data are brushed out into carry out persistence from cpu cache using clflush and mfence, Then carry out original place to data page again writes renewal, makes the write performance for improving small data in this way, in system crash or disconnected After electricity, the data read in small data daily record carry out redo, so as to complete the operation do not submitted, prevent that data are inconsistent State.

5. a kind of small letter optimization method of JFS based on nonvolatile memory according to claim 1, it is special Sign is that, in the 7th described step, it is specially：

It is related to multiple pages when judging that data update, the data of change part is write in newly assigned space, then the original do not changed For data duplication to the new page, more new metadata and daily record entry point to the pointer of new page, and Update log tail carries out this The submission of secondary renewal, ensure system once being collapsed in reproducting periods, still can basis as long as metadata daily record tail is not submitted Old metadata finds original data page, so as to ensure that data are in the state of uniformity.