Still confused by LSI / AvagoTech Performance Recommendations

[Wixner]

Hi,

I'm just about to leave my old HDD array for a Consumed-Grade SSD array but there are some recommendations from LSI / AvagoTech that I'm still confused about: namely Write Cache Policy and Data Placement Policy.

This is from their "Benchmark_Tips" document (LSI MegaRAID Controller Benchmark Tips (1891 KB) regarding SSD Performance Testing: For transactional I/O Benchmarking (My array will be a DeDuped' Hyper-V Storage array) RAID Write (presumably WriteBack?) is disabled and the Data Placement is set to Direct IO. This is also the requirements for their Fast Path technology.

Using this configuration will bypass the Controller Cache entirely even though it is faster (512MB DDR2 800MHz has a theoretical bandwith of 6400MB/s) than the theoretical speed of the Dual Link to the backplane (6GBps * (Multilane, 4) * (Ports, 2) = 48GBps / 8 = 6GB/s)

This configuration confuses med even more should i change controller to a 9270-8i-based one with 1GB 1333MHz DDR3 Cache that can peak 10.6GB/s.

I know this is all theoretical but what am I missing? Why should i write directly to the SSD's instead of the Cache? BBU and CacheCade 2.0 / FastPath Key is available.

 

[Jesin A]

I asked a somewhat similar question on ServerFault - How is LSI FastPath different from Software RAID?

From my understanding it comes down to latency, when write back is enabled the controller has more work to do managing data between its cache and the disks. The controllers are fast enough to do this with mechanical drives. As SSDs already operate at low latencies, these write back operations create an additional overhead thus defeating the purpose of a caching.

 

[Patriot]

I asked a somewhat similar question on ServerFault - How is LSI FastPath different from Software RAID?

From my understanding it comes down to latency, when write back is enabled the controller has more work to do managing data between its cache and the disks. The controllers are fast enough to do this with mechanical drives. As SSDs already operate at low latencies, these write back operations create an additional overhead thus defeating the purpose of a caching.

Also... if you have plp SSDs ... you can turn on the local SSD cache... no need for controller cache... it just adds latency.

 

[Wixner]

It makes sense if it is all about latency in the cache.
So "Direct-IO" with "Write-Through" and "No Read Ahead" is the way to go with SSD-only VD's.

Any thoughts on the "Stripe-Size"? The recommendation is 64KB and i know that it as been a huge debate on this topic but that is for mechanical drives only.