BUFFER HASH KV TABLE

Transcription

1 BUFFER HASH KV TABLE CHEAP AND LARGE CAMS FOR HIGH PERFORMANCE DATA-INTENSIVE NETWORKED SYSTEMS PAPER BY ASHOK ANAND, CHITRA MUTHUKRISHNAN, STEVEN KAPPES, ADITYA AKELLA AND SUMAN NATH PRESENTED BY PRAMOD KONDRU

2 OUTLINE INTRODUCTION MOTIVATION CHALLENGES FLASH STORAGE AND HASH TABLES BUFFER HASH DATA STRUCTURE SUPER TABLE SUPER TABLE OPERATION INCARNATION EVICTION

3 INTRODUCTION A number of data-intensive networked systems have emerged where there is a need to maintain hash tables as large as tens to a few hundred gigabytes in size. Example: WAN optimizers. The key requirement is that the mechanisms used be cost-effective for the functionality they support i.e. the mechanisms should offer a high number of hash table operations (> 10K) per second while keeping the overall cost low. The mechanisms that satisfy these requirements are referred to as CLAMs (cheap and large CAMs).

4 MOTIVATION To design and evaluate an approach that is 1-2 orders of magnitude better in terms of hash operations/sec/$ compared to both disk-based and DRAM based approaches. This approach uses a commodity two-level storage/memory hierarchy consisting of some DRAM and a much larger amount of flash storage. The design consumes most of the I/Os in the DRAM, giving low latency and high throughput I/Os compared to a flash-only design.

5 ADDRESING CHALLENGES A key idea behind Buffer Hash is that instead of performing individual random insertions directly on flash, DRAM can be used to buffer multiple insertions and writes to flash can happen in a batch. This shares the cost of a flash I/O operation across multiple hash table operations, resulting in a better amortized cost per operation.

6 CONTINUED.. FAST LOOKUP: To reduce the overhead of examining on-flash batches, Buffer Hash 1. Partitions the key space to limit the lookup to one partition, instead of the entire flash 2. Uses in-memory Bloom filters (as Hyperion does) to efficiently determine a small set of batches that may contain the key. LIMITED FLASH: Buffer Hash uses a novel age-based internal organization that naturally supports bulk evictions of old entries in an I/O-efficient manner.

7 CONTINUED.. PERFORMANCE TUNING: UPDATES: Parameters involved in design of CLAMs, such as the amount of DRAM to use, and the sizes of batches and Bloom filters. To support good update latencies, we adopt a lazy update approach where all value mappings, including deleted or updated ones, are temporarily left on flash and later deleted in batch during eviction.

8 QUESTION 1 A key idea behind BufferHash is that instead of performing individual random insertions directly on flash, DRAM can be used to buffer multiple insertions and writes to flash can happen in a batch. Very briefly explain the difference between the ways of FAWN and BufferHash in which they locate a KV pair written on the flash? The KV pair is located directly in the flash in a random manner whereas the KV pair is first looked up in the buffer using a Bloom Filter and then the bloom filter again is responsible for KV pairs in the Flash Memory.

9 FLASH STORAGE AND HASH TABLES Applications should avoid random writes, in-place updates, and sub-block deletions as they are significantly expensive on flash. Since reads and writes happen at the granularity of a flash page (or an SSD sector), an I/O of size smaller than a flash page (2KB) costs at least as much as a full-page I/O. Thus, applications should avoid small I/Os if possible. The high fixed initialization cost of an I/O can be amortized with a large I/O size. Thus, applications should batch I/Os whenever possible.

10 BUFFER HASH DATA STRUCTURE

11 DATA STRUCTURE To allow multiple insertions to be performed all at once, BufferHash operates in a lazy batched manner: it accumulates insertions in small in-memory hash tab When a buffer fills up, all inserted items are pushed into flash in a batch.les (called buffers), without actually performing the insertions on flash. For I/O efficiency, items pushed from a buffer to flash are sequentially written as a new hash table, instead of performing expensive update to existing in-flash hash tables. Thus, at any point of time, the flash contains a large number of small hash tables.

12 QUESTION 2 BufferHash consists of multiple super tables. Each super table has three main components: a buffer, an incarnation table, and a set of Bloom filters. Use Figure 1 to describe BufferHash s data structure.

13 SUPER TABLE Buffer Hash consists of multiple super tables. Each super table has three main components: a buffer, an incarnation table, and a set of Bloom filters. These components are split into two level hierarchy:-components in the higher level are maintained in DRAM, while those in the lower level are maintained in flash.

14 SUPER TABLE CONTD.. BUFFER: This is an in-memory hash table where all newly inserted hash values are stored. INCARNATION TABLE: The buffers flushed to flash are called incarnations. This is an in-flash table that contains old and flushed incarnations of the in-memory buffer. BLOOM FILTER: To avoid the excessive I/O cost, a super table maintains a set of in-memory Bloom Filters, one per incarnation.

15 QUESTION 3 This is an in-flash table that contains old and flushed incarnations of the inmemory buffer. Please explain the relationship between the buffer and the incarnation. Buffer is the collection of KV pairs in the Hash Table present on the DRAM. Incarnation is the buffers which are moved into the Flash memory when the buffer gets filled.

16 QUESTION 4 Since the incarnation table contains a sequence of incarnations, the value for a given hash key may reside in any of the incarnations depending on its insertion time. Please explain why Bloom filters are needed. Since the Hash Tables are updated in batches, KV pair lookups will be very expensive on flash. But by using a bloom filter for each batch makes lookups faster and less expensive.

17 QUESTION 5 A super table supports all standard hash table operations Describe the steps involved in insert, lookup, update/delete operations.

18 SUPER TABLE OPERATIONS INSERT: To insert a (key, value) pair, the value is inserted in the hash table in the buffer. If the buffer does not have space to accommodate the key, the buffer is flushed and written as a new incarnation in the incarnation table.

19 SUPER TABLE OPERATIONS LOOKUP: A key is first looked up in the buffer. If found, the corresponding value is returned. Otherwise, in-flash incarnations are examined in the order of their age until the key is found. Bloom filters are used to check for in-flash lookups.

20 SUPER TABLE OPERATIONS UPDATE/DELETE: Flash does not support small updates/deletions efficiently; hence, we support them in a lazy manner.

21 INCARNATION EVICTION There are two Buffer Hash Primitives: 1. Full Discard: Evict all items(fifo). 2. Partial Discard: Retain few items based on application policy Buffer Hash best suited for FIFO Incarnations arranged by age Other useful policies at some additional cost

22 QUESTION 6 If the Bloom filter matches, the incarnation is read from flash, and checked if it really contains the key. Note that since each incarnation is in fact a hash table,.. Could you describe the structure of the hash table? Could we hold all incarnations hash tables in the memory? Why?

23 QUESTIONS