StarWind Software

I'm using Starwinds Free Edition for our development environment, and have come across some issues with performance. We're using the following hardware and corresponding configurations below. Was wondering if anything was out of the ordinary or not to some specification that is tried and true.

Starwind iSCSI Host:
Dell R710
- Dual 2.4 Ghz E5620 Processors
- 36 GB RAM
- Windows 2008 R2
- Storage
- Local storage
- Dell PERC 6i
- ~1TB RAID 10
- 64k Allocation Size
- Device type Deduplicated Disk
- 500GB
- Deduplication Cache Size 10MB
- Metadata Cache Size 1002MB (Automatic)
- Deduplication block size 4k
- No Caching
- Dell PERC 5E external Controller
- Dell MD3000 Direct Attached Storage
- ~1TB RAID 5 vDisk
- Segment Size 256kb
- 64k Allocation Size
- Device type Image File
- Header Size 0
- Asynchronous Mode Yes
- No Caching
- Network
- LAN
- Intel Gigabit ET Quad Server Adapter Team
- Connected to server LAN
- Segregated from iSCSI network on 10.x.x.x subnet
- iSCSI
- Broadcom NetExtreme II NIC Team
- Smart Load Balancing
- Enabled Team Offload Capabilities TOE, LSO, CO, RSS
- Team MTU 9000
- All 4 ports connected to dedicated Cisco Switch
- Jumbo Packets enabled
- Segregated IP local to switch on 192.x.x.x network

VMware Hosts
Dell R710
- iSCSI networking
- Broadcom NetExtreme II NIC
- Software HBA
- MTU 9000
- 192.x.x.x IP Address

Cisco 2970 Switch
- Jumbo Packets enabled
- 192.x.x.x IP; not connected to LAN

---

Symptoms:
- Very slow performance of VMs stored on Starwinds iSCSI
- Instability of hosts connected to iSCSI network
- Very high latency on VMware datastores located on iSCSI

---

Are there any comprehensive guides available that show configuration and use of the Starwinds solution from end-to-end?

Pretty much everything is broken...

1) Windows 2008 R2 should be replaced with 2012 R2 as we do use performance-related features from it (TCP loopback accelerator, light-weight user-mode threads, new synchronization objects, advanced memory management etc)

2) Running StarWind with just 10MB of dedupe cache and no upper-level cache is insane... Performance would just suck. Use automatic settings for LSFS internal dedupe cache size and allocate ±10% of volume size RAM cache. It's recommended to upgrade for HA as it has safe write-back cache (running write-back RAM cache on a single controller is dangerous) and also HA has multiple I/O paths.

3) Configuring FLAT on top of a parity RAID5 is by far the best idea. Parity RAIDs to read-modify-write so every single 4KB-8KB write would trigger whole 256KB+ parity stripe update. Combine this with your lack of cache (disabled) and I guess that's the main reason why your performance is not great. Use RAID10 with VM workloads or use LSFS as you CAN use it with parity RAIDs and RAID0 (it does log-structuring). Get latest RC2 to play with LSFS.

4) You should never team iSCSI connections. In the best case only one physical NIC would work as by default you don't configure MC/S (and it really never pays back when configured unless in tests). Unteam your NICs and use MPIO in Round-Robin mode. That's the way to go!

5) ... That's to start with. Please fix these issues and we'll continue. Also make sure you have raw disk and raw network (TCP) performance numbers when talking about StarWind performance tuning as if we don't have basic numbers we have no clue how fast we can go.

Hope this helped

anton (staff) wrote:Unteam your NICs and use MPIO in Round-Robin mode. That's the way to go!

Why does the guide show to use failover only instead of round robin?

http://www.starwindsoftware.com/styles- ... manual.pdf

I guess it's because:

1) V8 and not V6

2) Hyper-Converged instead of a Compute and Storage separated

3) Hyper-V and not vSphere

rrbnc wrote:

anton (staff) wrote:Unteam your NICs and use MPIO in Round-Robin mode. That's the way to go!

Why does the guide show to use failover only instead of round robin?

http://www.starwindsoftware.com/styles- ... manual.pdf