VLDB2024_Optimizing Distributed Tiered Data Storage Systems with DITIS_华为.pdf

迹部景吾

615

4页

8次

2024-09-09

免费下载

Optimizing Distributed Tiered Data Storage Systems with DITIS

Sotiris Vasileiadis

Cyprus University of Technology

Limassol, Cyprus

sr.vasileiadis@edu.cut.ac.cy

Matthew Paraskeva

Cyprus University of Technology

Limassol, Cyprus

mp.paraskeva@edu.cut.ac.cy

George Savva

Cyprus University of Technology

Limassol, Cyprus

gec.savva@edu.cut.ac.cy

Andreas Efstathiou

Cyprus University of Technology

Limassol, Cyprus

andreasefstathiouudt@gmail.com

Edson Ramiro Lucas Filho

Cyprus University of Technology

Limassol, Cyprus

edson.lucas@cut.ac.cy

Jianqiang Shen

Huawei Technologies Co., Ltd.

Shenzhen, China

shenjianqiang@huawei.com

Lun Yang

Huawei Technologies Co., Ltd.

Shenzhen, China

yanglun12@huawei.com

Kebo Fu

Huawei Technologies Co., Ltd.

Shenzhen, China

fukebo@huawei.com

Herodotos Herodotou

Cyprus University of Technology

Limassol, Cyprus

herodotos.herodotou@cut.ac.cy

ABSTRACT

Modern data storage systems are characterized by a distributed

architecture as well as the presence of multiple storage tiers and

caches. Both system developers and operators are challenged with

the complexity of such systems as it is hard to evaluate how a con-

guration change will impact the workload or system performance

and identify the best conguration to satisfy some performance

objective. DITIS is a new simulator that models the end-to-end

execution of le requests on distributed tiered storage systems that

addresses the aforementioned challenges eciently without any

costly system redeployments. The demonstration will showcase

the key functionalities and benets oered by DITIS, including

(i) analyzing workload traces to understand their characteristics

and the behavior of the underlying storage system; (ii) running

simulations with dierent congurations to evaluate their impact

on performance; and (iii) running optimizations over custom search

spaces to nd the best conguration that satises a given objective.

PVLDB Reference Format:

Sotiris Vasileiadis, Matthew Paraskeva, George Savva, Andreas Efstathiou,

Edson Ramiro Lucas Filho, Jianqiang Shen, Lun Yang, Kebo Fu,

and Herodotos Herodotou. Optimizing Distributed Tiered Data Storage

Systems with DITIS. PVLDB, 17(12): 4393 - 4396, 2024.

doi:10.14778/3685800.3685883

PVLDB Artifact Availability:

The source code, data, and/or other artifacts have been made available at

https://github.com/cut-dicl/ditis-ui.

1 INTRODUCTION

Modern data storage systems exhibit considerable complexity due

to their distributed nature and the need to balance data and load

This work is licensed under the Creative Commons BY-NC-ND 4.0 International

License. Visit https://creativecommons.org/licenses/by-nc-nd/4.0/ to view a copy of

this license. For any use beyond those covered by this license, obtain permission by

emailing info@vldb.org. Copyright is held by the owner/author(s). Publication rights

licensed to the VLDB Endowment.

Proceedings of the VLDB Endowment, Vol. 17, No. 12 ISSN 2150-8097.

doi:10.14778/3685800.3685883

across the storage nodes [

]. In addition, these systems incorporate

multiple storage tiers, encompassing numerous HDDs and SSDs,

along with multiple cache levels of DRAM and NVRAM. Conse-

quently, new data management policies are required to optimize

performance and resource utilization. Furthermore, these systems

integrate diverse redundancy mechanisms, including replication

and erasure coding, to ensure data durability and fault tolerance.

The multifaceted architecture of modern data storage systems

necessitates sophisticated management strategies to harness their

full potential, for both system developers and operators. For devel-

opers, evaluating the impact of new policies for caching, tiering,

and other mechanisms is cumbersome and time-consuming as it

requires system redeployments. Hence, it is very dicult to explore

the design space for promoting changes in the system. For opera-

tors, it is challenging to evaluate how their workloads will behave

after a system reconguration or upgrade as well as determine the

best system conguration that will satisfy their objectives.

Simulation presents a logical approach to address the aforemen-

tioned challenges. Numerous simulators concentrate on modeling

particular aspects of storage systems, including caching and tier-

ing policies [

], scheduling [

], network communication [

], and

le system behavior [

]. Some simulators are also available for

simulating either single-node multi-tier storage systems [4, 10] or

distributed single-tier storage systems [

]. However, none of

the current simulators can fully encompass the complexity and

nuances of contemporary storage systems that feature multiple

storage nodes, diverse storage tiers, and various cache levels.

DITIS [

] is a new comprehensive simulator that models the end-

to-end execution of le requests on distributed multi-tier storage

systems. The key novelties of DITIS include (i) an architecture based

on an adaptation of the actor model instead of the typical event-

oriented or process-oriented models; (ii) a machine learning-based

initialization process for placing data to the appropriate tier/cache

before the simulation begins; and (iii) ne-grained but ecient

performance cost models for HDD, SSD, NVRAM, DRAM, and net-

work communications. Moreover, DITIS is extremely congurable

with 131 conguration parameters touching all aspects of a storage

system (e.g., number of nodes/tiers/caches, device and network

4393

Storage NodeStorage Node

Access Layer

Persistence Layer

Workload Replay^*

Application

Application^

Workload Initializer^*

Trace ParserTrace Parser

File Home Layer

Persistence Module^

Redundancy

Policy*

Block

Balancer*

File Home Module^

Metadata

Manager

Storage Pool

Manager

Cold Storage Pool

SATA/NL-SAS SSDs or HDDs

Warm Storage Pool

SAS SSDs

Hot Storage Pool

NVMe SSDs

Access Module^

Metadata

Manager

Cache

Dataflow

Manager

Dataflow

Policies*

Dataflow

Manager

Dataflow

Policies*

Dataflow

Manager

Dataflow

Policies*

Cache

Manager

L2 Cache

Policies*

Cache

Manager

L0 Cache

Policies*

L0 Data

Cache

Manager

L1 Cache

Policies*

L1 Data

Cache

Tiering

Manager

Tiering

Policies*

L2 Data Cache

File Home

Module^

Access

Module^

Persistence

Module^

Application

Connector*

File

Balancer*

Figure 1: DITIS architecture. Components marked with ^ are

actors and with

∗

are pluggable policies.

characteristics, etc.) as well as extensible with 44 pluggable policies

controlling all aspects of data ow, caching, and tiering decisions.

This demonstration aims at showcasing the benets to both

storage system developers and operators from using DITIS to:

•

Understand workload characteristics and the behavior of the

underlying storage system;

•

Evaluate the impact of dierent storage conguration setups and

policies to the workload performance; and

•

Optimize the storage conguration to satisfy a workload or sys-

tem objective such as maximizing throughput or cache hit ratio.

2 DITIS SIMULATOR

DITIS is a simulator designed for distributed and tiered le-based

storage systems, capable of handling up to three storage tiers plus

three levels of caches. It allows for the conguration of various

storage media devices such as HDD, SSD, NVRAM, and DRAM, each

with specic performance characteristics. The simulator calculates

the duration of I/O requests for each tier, cache, and network using

detailed performance cost models for each device type.

2.1 System Architecture

DITIS requires two inputs to simulate a workload execution on

a storage system: (1) a text le containing a workload trace with

le requests containing details like process ID, timestamp, le op-

eration, oset, length, le size, and original duration; and (2) a

storage conguration that denes the storage system’s behavior

and structure, allowing for the customization of components of the

storage system, specifying their performance characteristics, and

determining which policies to use during the simulation. The main

system components of DITIS, shown in Figure 1, are:

•

Workload Initializer: Sets up the initial state of the system

(e.g., creates pre-existing les) before the trace is executed.

•

Workload Replay: Controls the order and timing of submitting

and simulating the le requests from the workload trace.

•

Application: Represents external applications (processes) that

submit le requests to the storage system in parallel.

•

Application Connector: Balances incoming application con-

nections to the available Access Modules.

•

Access Module: Represents a storage system’s access compo-

nent running on a storage node or a system’s client.

•

File Balancer: Distributes les and le requests to File Home

Modules based on full le paths.

•

File Home Module: Maintains a partition of the system’s names-

pace on a storage node.

•

Block Balancer: Distributes data blocks and block requests to

Persistence Modules based on a block IDs (or addresses).

•

Persistence Module: Stores and processes data blocks in tiered

storage pools on a storage node.

•

Redundancy Policy: Controls the redundancy policy of blocks

(e.g., using erasure coding, replication, or RAID).

DITIS generates an output trace, accurately capturing the sequence

of le requests from the input trace but with a simulated duration

for each operation. Alongside, DITIS produces a detailed report

containing information and statistics about the workload execution

and the storage system (described in Section 3).

2.2 Simulation Process

Following the Actor Model, the key components of the system are

treated as actors (marked with ^ in Figure 1). These actors main-

tain their own private state, process messages received from other

actors, and send messages to other actors enabling asynchronous

message exchange, while respecting a simulation clock. This ap-

proach ensures the sequence and timing of events accurately reect

those of a real distributed multi-threaded system. In DITIS, the

process for handling I/O requests begins when an Application con-

nects to an Access Module and sends a le request (e.g., a read

request) to the storage system. If the request can be served by the

Access Module (e.g., perform the read from the local cache), a re-

spond is sent back to the Application. Otherwise, the request is

forwarded to the appropriate File Home Module for processing.

The request can be completed there or forwarded as one or more

block requests to the appropriate Persistence Module(s). There, the

request is served from the appropriate storage pool and returned

upstream. While the requests are getting processed, all relevant

decisions are made by pluggable policies, such as cache admission,

eviction, and prefetching, tier migrations, block redundancy, etc.

At each step, the duration is calculated based on several pluggable

performance cost models, while taking into account the potential

concurrent execution of requests as well as resource contention.

2.3 Performance Cost Modeling

The performance cost models for all devices and network are de-

signed to be pluggable, thus making it easy to replace to accommo-

date dierent simulation needs or advancements in technology.

4394