1HEC FSIO 2009

Parallel File System Simulator In order to test the Parallel File System (PFS)

scheduling algorithms in a light-weighted approach, we have developed the PFS simulator.

We use discrete event simulator and accurate disk modeling software to construct an agile parallel file system simulator.

This simulator is capable of simulating enough details while having an acceptable simulation time.

tests. tests.

2HEC FSIO 2009

Existing PFS simulators The IMPIOUS simulator, by E Molina-Estolano, al.

et[1]. It does not model the metadata server. The scheduler modules are lacking, so scheduling

algorithms are hard to model. The simulator developed in PVFS improvement paper

by Carns P. H., al. et[2]. It over simulates the network, which extends the

simulation time. It uses real PVFS in simulation, which introduces too much

details, while not flexible.

3HEC FSIO 2009

Simulator Details We use the discrete event simulation library

OMNeT++ 4.0 to simulate the network. It is capable of simulating the network topology with

bandwidth and delay. We use Disksim to simulate the data server disks.

Disksim accurately estimates the time for data transactions on the physical disks.

Disksim allows users to extract disk characteristics from real disks.

4HEC FSIO 2009

Simulator Architecture

Client

trace

MetadataServer

StrippingStrategy

Client

trace

Client

trace

Simulated Network

OMNeT++

Disksims instances

Data ServersSchedulingAlgorithm

Socket Connections

MetadataServer

Local FS

Diskqueue

Local FS

Disk

queue

Local FS

Disk

queueoutput

New request arrives.

Query file layout.

Send the request to data server.

Dispatch the job.

Send the results back.

5HEC FSIO 2009

Implementation of Scheduling Algorithms

Similarly, we have also implemented FIFO, Distributed SFQ[3], MinSFQ[4] algorithms, etc.

/* SFQ algorithm, at each data server */systime = 0waitQ.initiate()while(!simulation_end) { if reqArrive(), then: R = getReq() R.start_tag = min { R.getPrevReq().finish_tag, systime } R.finish_tag = R.start_tag + R.cost / R.getFlow().weight pushReq(R, waitQ) if diskHasSlot(), then: R = popReqwithMinStartTag(waitQ) systime = R.start_tag dispatch(R)}

The request from client arrives at the data server.

Disksim tells OMNet++ that it still has free slot.

OMNet++ dispatches the request to Disksim.

6HEC FSIO 2009

Simulation 1: Weighted Scheduling 2 groups of clients, 16 each. 4 Data servers and 1 Metadata server. Parallel Virtual File System(v2) is modeled. All files striped to all servers. Stripe size: 256KB. The trace files are generated by IOR.

Each client has 100MB checkpoint-write operation. SFQ with different weight assignments have been

simulated. The simulated weight ratios are 1:1, 2:1 and 10:1.

7HEC FSIO 2009

Throughput ResultsFIFO

010000200003000040000500006000070000

1 3 5 7 9 11 13 15 17 19 21 23 25

Time (s)

Thro

ughp

ut(b

yte/

s)

Group 1

Group 2

SFQ (1:1)

010000200003000040000500006000070000

1 3 5 7 9 11 13 15 17 19 21 23 25

Time (s)

Thro

ughp

ut(b

yte/

s)

Group 1

Group 2

SFQ (2:1)

0

20000

40000

60000

80000

100000

1 3 5 7 9 11 13 15 17 19

Time (s)

Thro

ughp

ut(b

yte/

s)

Group 1

Group 2SFQ (10:1)

0

20000

40000

60000

80000

100000

120000

1 3 5 7 9 11 13

Time (s)

Thro

ughp

ut(b

yte/

s)

Group 1

Group 2

8HEC FSIO 2009

Simulation 2: Scheduling Fairness Same as Simulation 1, except that the two groups do

not have the same resources: Group 1’s files are stripped to all servers [1, 2, 3, 4]. Group 2’s files are stripped to 3 servers [1, 2, 3].

We add Distributed SFQ algorithm[3], in which all servers share the scheduling information.

We are able to show that if the workloads are not evenly distributed, the Distributed SFQ algorithm achieves better fairness than FIFO and SFQ.

9HEC FSIO 2009

Throughput Results

FIFO

010000200003000040000500006000070000

1 3 5 7 9 11 13 15 17 19 21 23 25

Time (s)

Thro

ughp

ut(b

yte/

s)

Group 1

Group 2

SFQ (1:1)

010000200003000040000500006000070000

1 3 5 7 9 11 13 15 17 19 21 23 25

Time (s)

Thro

ughp

ut(b

yte/

s)

Group 1

Group 2

DSFQ (1:1)

010000200003000040000500006000070000

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Time (s)

Thro

ughp

ut(b

yte/

s)

Group 1

Group 2

10HEC FSIO 2009

References[1] E Molina-Estolano, C Maltzahn, J Bent and S A Brandt, “Building a parallel file system

simulator”, Journal of Physics: Conference Series 180 (2009) 012050.[2] Carns P H, Ligon W B, al. et. “Using Server-to-Server Communication in Parallel File Systems to

Simplify Consistency and Improve Performance”, Proceedings of the 4th Annual Linux Showcase and Conference (Atlanta, GA) pp 317-327.

[3] Yin Wang and Arif Merchant, “Proportional Share Scheduling for Distributed Storage Systems”, File and Storage Technologies (FAST’07), San Jose, CA, February 2007.

[4] W. Jin, J. S. Chase and J. Kaur, “Interposed Proportional Sharing For A Storage Service Utility”, SIGMETRICS, E. G. C. Jr., Z. Liu, and A. Merchant, Eds. ACM, 2004, pp. 37-48.