现代RDMA网络下的数据库高可用性重新思考

版权申诉

56 浏览量更新于2024-08-13 收藏 690KB PDF 举报

"这篇PDF论文探讨了在使用RDMA（远程直接内存访问）网络的背景下重新思考数据库高可用性的问题。作者包括Erfan Zamanian、Xiangyao Yu、Michael Stonebraker和Tim Kraska，分别来自布朗大学和麻省理工学院。他们提出，传统的主动-被动和主动-主动复制算法在网络性能是主要瓶颈的时代设计，但随着新一代高吞吐量、低延迟网络（如RDMA网络）的出现，这种假设需要重新评估。" 正文: 随着技术的发展，数据库系统的高可用性已经成为关键问题，尤其是在处理机器故障时。传统的数据复制方法，如主动-被动和主动-主动复制策略，都是在假设网络是性能瓶颈的情况下设计的。这些方法主要关注减少副本间的网络通信，以换取更多的处理冗余，这在分布式数据库设计的传统智慧中是有道理的。然而，《Rethinking Database High Availability with RDMA Networks》论文指出，随着RDMA网络的出现，这一情况发生了变化。RDMA允许数据直接在内存之间传输，减少了CPU的干预，因此现在CPU可能成为新的性能瓶颈。现有的网络优化复制技术在这样的环境下不再是最优选择。论文提出了“Active-Memory Replication”（活动内存复制）作为一种新的高可用性方案，它旨在利用RDMA网络的优势，减少CPU的负载，并优化数据复制过程。这种方法可能更有效地适应现代网络环境，提供更高的数据可用性和系统性能。在高可用性数据库系统中，快速的数据恢复和一致性保证是至关重要的。传统的复制技术可能在处理大量数据传输和快速故障切换时面临挑战，而RDMA网络的低延迟特性为这些问题提供了可能的解决方案。通过将内存访问的责任转移到网络，Active-Memory Replication可能会降低系统延迟，提高吞吐量，同时保持数据的一致性。此外，论文还可能讨论了如何在RDMA网络中实现Active-Memory Replication的具体技术细节，如如何协调多个副本，如何处理网络故障，以及如何在不增加过多复杂性的情况下确保数据安全性。这些内容对于理解如何在现代数据中心环境中构建高效且可靠的数据库系统至关重要。这篇论文挑战了传统观念，指出在RDMA网络环境下，数据库高可用性的实现需要对现有复制策略进行根本性的反思。通过引入Active-Memory Replication，作者提供了一种新的思路，以适应网络技术的进步，进一步推动数据库系统性能的提升。

Exec%T1

Commit

Exec%T2%(A)

Exec%T2%(B)

Commit

Exec%T1

Commit

Exec%T2%(B)

Exec%T2%(A)

Commit

Figure 2: Active-active replication in H-store/VoltDB

all replicas of the database in the same deterministic order, such

that all the replicas end up with identical states once the batch is

executed. Replicas in an active-active database only coordinate to

determine the batches, but do not coordinate during the execution

of transactions, which signiﬁcantly reduces the network trafﬁc be-

tween replicas. Two prominent examples of databases which use

active-active replication are H-store [16] (and its commercial suc-

cessor VoltDB [35]) and Calvin [37].

H-Store’s replication protocol is illustrated in Figure 2. In H-

Store, all transactions have to be registered in advance as stored

procedures. The system is optimized to execute each transaction

from the beginning to completion with minimum overhead. In H-

Store, transactions do not get record locks, and instead only lock

the partition they need. Transactions are executed in each parti-

tion sequentially, without getting pre-empted by other concurrent

transactions. For a single-partition transaction such as T 1, the pri-

mary replicates the ID of the invoked stored procedure along with

its parameters to all its backup replicas. All replicas, including the

primary, start executing the transaction code in parallel. Unlike in

log shipping, here the replicas do not need to coordinate, as they ex-

ecute the same sequence of transactions and make deterministic de-

cisions (commit or abort) for each transaction. For multi-partition

transactions, one of the primaries acts as the transaction coordina-

tor and sends the stored procedure and its parameters to the other

participating partitions. At each partition, an exclusive lock is ac-

quired, so that no other transaction is allowed to be executed on that

partition. Each partition sends the stored procedure to all its backup

replicas so they run the same transaction and build their write-set.

Finally, the coordinator initiates a 2PC to ensure that all the other

primaries are able to commit. H-Store performs extremely well if

the workload consists of mostly single-partition transactions. How-

ever, its performance quickly degrades in the presence of multi-

partition transactions, since all participating partitions are blocked

for the entire duration of such transactions.

Calvin [37] is another active-active system that takes a differ-

ent approach than H-Store to enforce determinism in execution and

replication (Figure 3). All transactions ﬁrst arrive at the sequencer

which orders all the incoming transactions in one single serial his-

tory. The inputs of the transactions are logged and replicated to all

the replicas. Then, a single lock manager thread in each partition

scans the serial history generated by the sequencer and acquires all

the locks for each transaction, if possible. If the lock is already

held, the transaction has to be queued for that lock. Therefore,

Calvin requires that the read-set and write-set of transactions are

known upfront, so that the lock manager would know what locks

to get before even executing the transaction (This assumption may

be too strict for a large category of workloads, where the set of

records that a transaction is read or modiﬁed is known throughout

executing the transaction). Those transactions which all their locks

are acquired by the lock manager are then executed by the worker

Sequencer

Sequencing

…

$%&'(!"()*+

$%&'(!#(),+

$%&'(!#()*+

-.//01

Figure 3: Active-active replication in Calvin

threads in each replica without any coordination between replicas.

For multi-partition transactions, the participating partitions com-

municate their results to each other in a push-based manner (instead

of pull-based, which is common in the other execution schemes).

Compared to Calvin with its sequencing overhead, H-store has

a much lower overhead for single-partitioned transactions. Calvin,

on the other hand, beneﬁts from its global sequencing for multi-

partition transactions.

3. THE CASE FOR REPLICATION WITH

RDMA NETWORKS

With the fast advancement of network technologies, conventional

log-shipping and active-active schemes are no longer the best ﬁts.

In this section, we revisit the design trade-offs that conventional

schemes made and demonstrate why the next-generation networks

call for a new design of high availability protocol in Section 3.1.

We then provide some background on RDMA in Section 3.2.

3.1 Bottleneck Analysis

The replication schemes described in the previous section were

designed in a time that network communication was the obvious

bottleneck in a distributed main-memory data store by a large mar-

gin. Reducing the need for accessing the network was therefore a

common principle in designing efﬁcient algorithms. Both classes

of techniques approach this design principle by exchanging high

network demand with more processing redundancy, each to a dif-

ferent degree. This idea is illustrated in Figure 4a. In log ship-

ping, the logs have to be replayed at each replica, which may not

be much cheaper than redoing the transaction itself for some work-

loads. Active-active techniques reduce the need for network com-

munication even further and thus improve performance when the

network is the bottleneck but impose even more redundancy for

computation.

In these networks, communication during replication is consid-

ered expensive mainly due to three factors. (1) Limited bandwidth

of these networks would be easily saturated and become the bottle-

neck. (2) The message processing overhead by the operating sys-

tem proved to be substantial [3], especially in the context of many

OLTP workloads which contain simple transactions that read and

modify only a few records. (3) High latency of network communi-

cation increases the transaction latency, contributing to contention

and therefore impacts throughput.

With the emergence of the next-generation of RDMA-enabled

networks, such as InﬁniBand, these assumptions need to be re-

evaluated. (1) Network bandwidth has increased signiﬁcantly, and

its increase rate does not seem to be slowing down [13]. For exam-

ple, a Mellanox ConnectX-4 EDR card offers 100× bandwidth of a

typical 1Gb/sec Ethernet found in many public cluster offerings (in-

cluding our own private cluster). (2) The RDMA feature open new

1639

剩余13页未读，继续阅读

工控老马

粉丝: 678
资源: 2561

现代RDMA网络下的数据库高可用性重新思考

Rethinking the Inception Architecture for Computer Vi.pdf

藏经阁-Rethinking Recurrent Neural Ne (1).pdf

RETHINKING THE VALUE OF NETWORK PRUNING.pdf

Rethinking Productivity in Software Engineering.pdf

rethinking_package.pdf

信息安全_数据安全_law-f03_rethinking_employee_surv.pdf

信息安全_数据安全_Virtuous Cycles：Rethinking the S.pdf

2020-Rethinking Pre-training and Self-training.pdf

gUnet源码(Rethinking Performance Gains in Image Dehazing Networks)

信息安全_数据安全_Rethinking_Access_Control_and_Au.pdf

最新资源