NetApp Vs. IBM SAN

[kriskros]

Hi,

I am looking to purchase a SAN solution to provide the storage for a blade centre. Initially starting with 2-3 blades attached to the SAN, all running SQL server and some Java app.
I've looked in detail at the IBM DS4300, done some testing and received some good results. However, am getting the hard sell on Netapp(probably FAS3020). Claims that iSCSI will surpass FC speeds due to channel bonding and that sort of thing. Certainly the copy features and interface on the Netapp look more modern. Anyone have any experience on the price/performance ratio? Additionally, the costing, reliability and general impressions?

 

[xmsre]

Well, it will - iSCSI will surpass FC speeds for SQL that is. It's not limited to Netapp. It doesn't need channel bonding even. For small block IO, the speed of iSCSI with jumbo frame support enabled on dedicated 1Ge infrastructure is significantly faster than 2G FCP. In FCP (FC4) there quite a bit of overhead for each IO transaction in the form of the FCP exchange and sequence. The larger your IO size, the better the throughput. For a 4K IO size, 1G FCP hits a max throughput of about 30MB/sec. That's not a Netapp number, EMC published it first in the paper "Top 10 Performance Optimizations for Exchange on the Symmetrix". Of course, as the IO size increases, the efficiency gets better. At a 64K IO size, the throughput is more like 80MB/sec on 1G FCP. The block size for a transactional application is driven by the page size of the application. SQL uses 8K pages. This isn't something you can change, it's something you design to. For iSCSI, there is only a thin sequence overhead. With jumbo frames, each frame can carry a payload up to 9000 bytes (compare to FCP frame of 2112 bytes). Observed speeds for small block IO on 1Ge iSCSI are 118MB/sec; very close to the theoretical maximum of 125MB/sec. There are hardware compression/accelleration products out there, and you do have multipath support over more than 1 NIC with MPIO.

Here's a good link from SNIA the compare the two:

http://www.snia.org/education/tutorials/spr2005/storage/#scsi-the-protocol

Now that we've talked about the channel, or how data gets to the storage, let's talk about the storage. When we talk about storage, the first thing is to understand the load. SQL has a couple different kinds of workloads, OLAP and OLTP. With OLAP, it's usually a read centric random workload against a data warehouse. For OLTP, it a more write centric random workload processing complex business transactions that may be spread across many tables. For a random read workload, caching provides some improvement and can be important. More important is to look for a predictive read ahead feature on the storage to keep the read cache full with relevant pages. For writes, it's important to understand the RAID type on the backend and the associated "write penalty". IBM DS is very traditional in this approach with RAID 5 and RAID 1/10/0+1. Netapp has a virtualiztion layer that:

1. Never overwrites
2. Coalesces writes.
3. Optimizes the order of write.

At the controller level, some vendors do 2 and 3 at the track level that can increase the performance of sequential writes. Netapp is the only one that does all three. The result is that there is no "write penalty" on Netapp. This can result in significanly fewer spindles for the same workload.

I would say do a side by side eval. I would use Iometer and the SNIA methodology for the test. You'll need to build workload profiles that represent the real workload of your SQL implementation:

The SNIA link for testing is:

http://www.snia.org/education/tutorials/fall2005/storage/#Storage-Performance-Testing

IBM positioning on this sounds a bit wierd, and is therefore suspect in my mind. They resell NetApp as the N series. Sounds like someone wants a bigger margin or has a quota to meet to me.