SPECsfs2008_nfs.v3 Result

Throughput (ops/sec) Response (msec) 158669 0.6 317906 0.7 477508 0.6 637151 0.7 797347 1.0 964648 1.1 1150533 1.0 1304474 1.4 1423859 2.0 1564404 2.2

Tested By	Avere Systems, Inc.
Product Name	FXT 3500 (44 Node Cluster)
Hardware Available	November 2011
Software Available	November 2011
Date Tested	October 2011
SFS License Number	9020
Licensee Locations	Pittsburgh, PA USA

The Avere Systems FXT 3500 appliance provides tiered NAS storage that allows performance to scale independently of capacity. The FXT 3500 is built on a 64-bit architecture that is managed by Avere OS software. FXT clusters scale to as many as 50 nodes and support millions of IO operations per second and deliver tens of GB/s of bandwidth. The Avere OS software dynamically organizes data into tiers. Active data is placed on the FXT appliances while inactive data is placed on slower mass storage servers. The FXT 3500 accelerates performance of read, write and metadata operations and supports both NFSv3 and CIFS protocols. The tested configuration consisted of (44) FXT 3500 cluster nodes. The system also included (4) OpenSolaris ZFS servers that acted as the mass storage systems. Avere's integrated global namespace functionality was used to present a single namespace to all clients.

Item No	Qty	Type	Vendor	Model/Name	Description
1	44	Storage Appliance	Avere Systems, Inc.	FXT 3500	Avere Systems Tiered NAS Appliance running Avere OS V2.1 software. Includes (15) 600 GB SAS Disks.
2	4	Server Chassis	Genstor	CSRM-4USM24ASRA	4U 24-Drive Chassis for OpenSolaris mass storage NFS Server.
3	4	Server Motherboard	Genstor	MOBO-SX8DTEF	Supermicro X8DTE Motherboard for OpenSolaris mass storage NFS Server.
4	8	CPU	Genstor	CPUI-X5620E	Intel Westmere Quad Core 2.4 GHz 12MB L3 CPU for OpenSolaris mass storage NFS Server..
5	8	CPU Heat Sink	Genstor	CPUI-2USMP0038P	SM 2U 3U P0038P Heat Sink for OpenSolaris mass storage NFS Server.
6	48	Memory	Arrow	MT36JSZF1G72PZ-1GD1	Micron DRAM Module DDR3 SDRAM 8GByte 240 RDIMM for OpenSolaris mass storage NFS Server.
7	92	Disk Drive	Hitachi	HDAS-3000H6G64U	3TB SATA II 64MB 3.5 ULTRASTAR 7200 RPM Disks for OpenSolaris mass storage NFS Server.
8	4	SSD	Intel	SSDSA2BZ200G301	Intel SSD 710 Series (200GB, 2.5in SATA 3 Gb/s, 25nm, MLC) for OpenSolaris mass storage NFS Server.
9	12	RAID Controller	Genstor	RAAD-LSAS9211-8i	LSI SAS 8 port PCI-X 6 Gbps LSI SAS 9211-8i HBA for OpenSolaris mass storage NFS Server.
10	24	Cabling	Genstor	CABL-SASICBL0281L	SAS 4-Lane for Backplane to 4-Lane SATA for OpenSolaris mass storage NFS Server.
11	4	Network Card	Intel	E10G42BTDA	Intel 10 Gigabit AF DA Dual Port Server Adapter for OpenSolaris mass storage NFS Server.

OS Name and Version	Avere OS V2.1
Other Software	Mass storage server runs OpenSolaris 5.11 svn_134, this package is available for download at http://genunix.org/dist/indiana/osol-dev-134-x86.iso
Filesystem Software	Avere OS V2.1

Name	Value	Description
zfs set atime	off	Disable atime updates on the OpenSolaris mass storage system.
ncsize	2097152	Size the directory name lookup cache (DNLC) to 2 million entries on the OpenSolaris mass storage system.
zfs:zfs_arc_max	0x1400000000	Set the maximum size of the ZFS ARC cache to 80GB on the OpenSolaris mass storage system.
zfs:zfs_arc_meta_limit	0x13c0000000	Limit the metadata portion of the ZFS ARC cache to 79GB on the OpenSolaris mass storage system.
zfs:zfs_vdev_cache_bshift	13	Limit the device-level prefetch to 8KB on the OpenSolaris mass storage system.
zfs:zfs_mdcomp_disable	1	Disable metadata compression on the OpenSolaris mass storage system.
zfs:zfs_txg_timeout	2	Aggressively push filesystem transactions from the ZFS intent log to storage on the OpenSolaris mass storage system.
zfs:zfs_no_write_throttle	1	Disable write throttle on the OpenSolaris mass storage system.
rpcmod:cotsmaxdupreqs	6144	Increase the size of the duplicate request cache that detects RPC-level retransmissions on connection-oriented transports on the OpenSolaris mass storage system.
ddi_msix_alloc_limit	4	Limit the number of MSI-X vectors per driver instance on the OpenSolaris mass storage system.
pcplusmp:apic_intr_policy	1	Distribute interrupts over CPUs in a round-robin fashion on the OpenSolaris mass storage system.
ixgbe.conf:mr_enable	0	Disable multiple send and receive queues for Intel 10GbE adapter on the OpenSolaris mass storage system.
ixgbe.conf:tx_ring_size	4096	Increase the packet receive ring to 4096 entries on the Intel 10GbE adapter on the OpenSolaris mass storage system.
ixgbe.conf:rx_ring_size	4096	Increase the packet transmit ring to 4096 entries on the Intel 10GbE adapter on the OpenSolaris mass storage system.
ixgbe.conf:intr_throttling	0	Disable interrupt throttling on the Intel 10GbE adapter on the OpenSolaris mass storage system.
ixgbe.conf:tx_copy_threshold	1024	Increase the transmit copy threshold from the default of 512 bytes to 1024 on the OpenSolaris mass storage system.
NFSD_SERVERS	1024	Run 1024 NFS server threads on the OpenSolaris mass storage system.
Writeback Time	12 hours	Files may be modified up to 12 hours before being written back to the mass storage system.
cfs.pagerFillRandomWindowSize	32	Increase the size of random IOs from disk.
cfs.quotaCacheMoveMax	10	Limit space balancing between cache policies.
cfs.resetOnIdle	true	Equally share cache capacity when system is not recycling any blocks.
vcm.readdir_readahead_mask	0x7ff0	Optimize readdir performance.
buf.autoTune	0	Statically size FXT memory caches.
buf.neededCleaners	16	Allow 16 buffer cleaner tasks to be active.
tokenmgrs.geoXYZ.fcrTokenSupported	no	Disable the use of full control read tokens.
tokenmgrs.geoXYZ.tkPageCleaningMaxCntThreshold	256	Optimize cluster data manager to write more pages in parallel.
tokenmgrs.geoXYZ.tkPageCleaningStrictOrder	no	Optimize cluster data manager to write pages in parallel.
cluster.VDiskCommonMaxRtime	600000000	Retry internal RPC calls for up to 600 seconds. Value is in microseconds.
cluster.VDiskCommonMaxRcnt	65535	Retry internal RPC calls for up to 65535 attempts.
cluster.VDiskInternodeMaxRtime	600000000	Retry inter-FXT node RPC calls for up to 600 seconds. Value is in microseconds.
cluster.VDiskInternodeMaxRcnt	65535	Retry inter-FXT node RPC calls for up to 65535 attempts.

Server Tuning Notes

21+1 RAID5 (RAIDZ) array was created on the OpenSolaris mass storage server using the following Solaris command: zpool create vol0 raidz ...list of 21 disk WWN's... log ...WWN of SSD device...

Description	Number of Disks	Usable Size
Each FXT 3500 node contains (15) 600 GB 10K RPM SAS disks. All FXT data resides on these disks.	660	342.2 TB
Each FXT 3500 node contains (1) 250 GB SATA disk. System disk.	44	10.0 TB
The mass storage systems contain (22) 3 TB SATA disks. The ZFS file system is used to manage these disks and the FXT nodes access them via NFSv3.	88	223.8 TB
The mass storage systems contain (1) 3 TB system disk.	4	10.9 TB
The mass storage systems contain (1) 200 GB SSD disk. ZFS Intent Log (ZIL) Device.	4	186.0 GB
Total	800	587.0 TB

Number of Filesystems	Single Namespace
Total Exported Capacity	229162 GB (OpenSolaris mass storage system capacity)
Filesystem Type	TFS (Tiered File System)
Filesystem Creation Options	Default on FXT nodes. OpenSolaris mass storage server ZFS filesystem created with 'zpool create vol0 raidz ...list of 21 disk WWN's... log ...WWN of SSD device...'
Filesystem Config	21+1 RAID5 configuration on OpenSolaris mass storage server.
Fileset Size	185647.1 GB

Item No	Network Type	Number of Ports Used	Notes
1	10 Gigabit Ethernet	44	One 10 Gigabit Ethernet port used for each FXT 3500 appliance.
2	10 Gigabit Ethernet	4	The mass storage systems are connected via 10 Gigabit Ethernet.

Network Configuration Notes

Each FXT 3500 was attached via a single 10 GbE port to one of two Arista Networks 7050-64 port 10 GbE switches. The FXT nodes were evenly split across the two switches. The load generating clients were also spread across the two switches. The mass storage servers were also attached to the same two switches, 2 mass storage servers per switch. Finally, the two switches were interconnected using a QSFP+ port along with 4x10GbE ports for a total of 80 Gbps of bidirectional inter-switch bandwidth. A 1500 byte MTU was used throughout the network.

An MTU size of 1500 was set for all connections to the switch. Each load generator was connected to the network via a single 10 GbE port. The SUT was configured with 88 separate IP addresses on one subnet. Each cluster node was connected via a 10 GbE NIC and was sponsoring 2 IP addresses.

Item No	Qty	Type	Description	Processing Function
1	88	CPU	Intel Xeon CPU E5620 2.40 GHz Quad-Core Processor	FXT 3500 Avere OS, Network, NFS/CIFS, Filesystem, Device Drivers
2	8	CPU	Intel Xeon E5620 2.40 GHz Quad-Core Processor	OpenSolaris mass storage systems

Processing Element Notes

Each file server has two physical processors.

Description	Size in GB	Number of Instances	Total GB	Nonvolatile
FXT System Memory	144	44	6336	V
Mass storage system memory	96	4	384	V
FXT NVRAM	2	44	88	NV
Grand Total Memory Gigabytes			6808

Memory Notes

Each FXT node has main memory that is used for the operating system and for caching filesystem data. A separate, battery-backed NVRAM module is used to provide stable storage for writes that have not yet been written to disk.

The Avere filesystem logs writes and metadata updates to the NVRAM module. Filesystem modifying NFS operations are not acknowledged until the data has been safely stored in NVRAM. The battery backing the NVRAM ensures that any uncommitted transactions persist for at least 72 hours. The OpenSolaris MASS contains a ZFS intent log (ZIL) device. The ZIL device is an Intel 710 SSD and this device will save all write-cached data in the event of a power loss. This SSD device is designed with sufficient capacitance to flush any committed write data to the non-volatile flash media in event of a power failure or unclean shutdown.

The system under test consisted of (44) Avere FXT 3500 nodes. Each node was attached to the network via 10 Gigabit Ethernet. Each FXT 3500 node contains (15) 600 GB SAS disks. The OpenSolaris mass storage systems were each attached to the network via a single 10 Gigabit Ethernet link. The mass storage servers were 4U Supermicro servers each configured with a software RAIDZ 21+1 RAID5 array consisting of (22) 3TB SATA disks. Additionally, a 200GB SSD device was used for the ZFS intent log (ZIL) in each mass storage server.

N/A

Item No	Qty	Vendor	Model/Name	Description
1	20	Supermicro	SYS-1026T-6RFT+	Supermicro Server with 48GB of RAM running CentOS 5.6 (Linux 2.6.18-238.19.1.el5)
2	2	Arista Networks	7050-64	Arista Networks 64 Port 10 GbE Switch. 48 SFP/SFP+ ports, 4 QSFP+ ports

LG Type Name	LG1
BOM Item #	1
Processor Name	Intel Xeon E5645 2.40GHz Quad-Core Processor
Processor Speed	2.40 GHz
Number of Processors (chips)	2
Number of Cores/Chip	6
Memory Size	48 GB
Operating System	CentOS 5.6 (Linux 2.6.18-238.19.1.el5)
Network Type	Intel Corporation 82599EB 10-Gigabit SFI/SFP+

Benchmark Parameters

Network Attached Storage Type	NFS V3
Number of Load Generators	20
Number of Processes per LG	176
Biod Max Read Setting	2
Biod Max Write Setting	2
Block Size	0

Testbed Configuration

LG No	LG Type	Network	Target Filesystems	Notes
1..20	LG1	1	/sfs0,/sfs1,/sfs2,/sfs3	LG1 nodes are evenly split across the two switches

Load Generator Configuration Notes

All clients were mounted against all filesystems on all FXT nodes.

Each load-generating client hosted 176 processes. The assignment of processes to network interfaces was done such that they were evenly divided across all network paths to the FXT appliances. The filesystem data was evenly distributed across all disks and FXT appliances and MASS storage servers.

N/A

• Config Diagram