On 02/18/2014 08:34 PM, Leonardo Santos wrote:
I'm trying to make a experimental balance striping between SSD and HDD
based on some performance parameter (latency or throughput) gotten on the
fly.
I realized that when I increase the number of operations for SSD (e.g. 3x1,
SSDxHDD) it doesn't give me the better of both.
My HDD read sequentially at ~100MB/s and my SSD at ~400MB/s. My hope is
throughput to be the sum of SSD and HDD, but this its not true for my tests.
For instance, when I put operations in 1x1:
SSD = ~100MB/s
HDD = ~100MB/s
When I put 3x1:
SSD = ~150MB/s
HDD = ~50MB/s (why its not 100MB/s?)
Not sure what you are trying to achieve and if you really need this or
just want to misuse DM to fit your goals.
1. Let's answer the "Why?" question
Are you trying to implement kind of I/O shaping or throttling?
(Otherwise I do not understand the need to cripple the SSD speed by
using it in such a configuration, but anyway...)
2. Then we can either move to "How?"
Would cgroups work for you?
For me its not parallel, but only distributes the operations in a
proportional way.
Questions:
- Why isn't throughput equal to HDD + SSD in normal mode (1x1)?
- Why HDD reduces proportionally?
- Its not parallel? Where can I change it to make this parallel?
3. Or to your "Why?" questions.
For that I have no answers, just questions:
- How do you measure the throughput? (Exact command would be the best.)
- How are the disks connected? (Is this single SSD you divided into
partitions and are now hitting it with 3x more IOPS while the reads are
not sequential anymore which _should_ not have much effect)
- What's the throughput when using striped SSD(s) only?
- What does `dmsetup status` says?
- What are sector sizes? Chunk sizes? and Block sizes you read?
I using this in my research and changed dm-stripe.c.
Thanks a lot!
*Leonardo Antônio dos Santos*
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
