Drive protection in SSD

[pshinde96]

Not sure if I understood the architecture correctly. I appreciate clarification on this.

I believe there is no RAID protection or hotspares for SSD. And also the SSD pool will have ONLY COPIES of hot data which can be write or read blocks. When we say copies, are these copies of blocks in the NVRAM or on the spindles?

Thanks

[Nick_Dyer]

Hello Prasad,

You are correct in saying that only copies of data resides in flash, which means no RAID or hotspares for them. The copies that are staged into flash are typically for randomly written data only (as sequentially written data has a high probability of being sequentially read back which doesn't need flash).

The overall write process is as follows:

1. Incoming data ingested to NVRAM, and is compressed inline as well as put into sequential stripes (we call this coalescing). At the same time we're also shadow copying this data into the NVRAM of the secondary controller.
2. Data is then shadow copied into DRAM.
3. Once stripes are complete data is sequentially written to NL-SAS, and hot blocks (random writes) are also inline copied to SSD on the fly.

As all data from NVRAM is written to DRAM and NL-SAS as part of the process you can say that we are copying blocks from all stages, but understanding the write process is key to understand the benefit of how CASL works under the hood.

[pshinde96]

Thanks Nick

I am also to trying to understand how the SSDs are pooled? Are they single SSDs or mutiple SSDs striped or concatenated?

For example in a 220-X4 the SSD capacity is 1.2TB. How many SSDs in this pool?

[dsandler71 · Accepted solution]

Prasad,

Each array has 4 SSDs in it. They make up the total SSD pool for that array. So, in your example, the CS220-X4 has 4 x 600GB SSDs for a combined pool of 1.2TB of RAW SSD (usable will be higher due to compression). We address each drive independently, that is to say we are not leveraging any type of RAID including RAID0. Think of it as a logical collection of completely independent drives. Each drive will receive its own completely independent data segment aligned to its block boundary. The next drive will receive another segment. So on and so forth. When you add an expansion shelf, that adds another SSD into the overall pool but the process remains the same.

As a side note, keep in mind that as you add capacity to expansion shelves there is an increase in the associated metadata. Hence, we don't consider the SSD in any expansion shelf as scaling the cache pool. The assumption is that although the overall pool size does increase, the available cache space will not increase significantly as an increase in metadata consumption of SSD will also occur.

[pshinde96]

Thank you all. What is the size of the NVRAM? Is this standard capacity in all models?

Where is the metadata stored?  NVRAM or both NVRAM and SSDs.

[aherbert23]

Prasad,

The NVRAM is 1GB in capacity but is never stressed to being full due to the way that CASL is able to sequentially drain the compressed data to disk so quickly. All data lives on disk (including the metadata) and is checksummed in self describing blocks that are protected by a dual parity RAID. This allows us to always provide accurate reliable reproduction of the data stored and protecting against disk bit flips or misdirected reads and writes. There is a statistical chance of these errors happening with all physical media thus our use of these protection mechanisms. Just as the flash is used as a cache for the data served it is also used as a cache for the metadata. Our SSD cache is never used as a single instance storage tier, but only as a read accelerating cache.

[rugby0134]

Here's a great video explaining the CASL process and how we use storage and NVRAM.

Nimble Storage CASL Architecture Technical Deep Dive - YouTube

The explanation starts at 24:05