Hello.
I have a several raids: RAID 0 SAS HDDs (each connected via 6Gb -- SAS-2). And 5+0 raid of SATA HDDs.
Both raids provide good read performance, but when I write to them, they start at insanely high speed (~1GB per second!). After some time, they become stuck for ~30 seconds. And then continue.
Write buffer is enabled.
Raids are from different vendors (Adaptec and IBM).
I believe they write data to the buffer and then need some time to flush it. It is reproducible only on huge (~ 20GB) files
Question 1: is it true? How can I profile/proof that fact?
I tried both: local access (copying from one raid to another and vice versa) and network access (via 10Gb network).
Effect is the same:
Users are angry when copying process is frozen, so my question is does it work as it should? If so, what can we do except migrating to SDDs (30TB of SSDs are expensive) or limiting my network speed?
We spent some time improving our network performance (by enabling RSS, jumbo frames, migrating to latest SMB version that supports multiple threads etc) and we now have ~9.7Gb/s with iperf3 between client and server.
Question 2: Is it possible to have at least 300/400 megabytes per second for writing with 10 or 50 on HDDs? Shouldn't we have ~150MB/s for each disk, 300MB for 2 HDDs in RAID0? Big files utilize sequental access, aren't they?
Question 3: RAIDs have huge stripe size (about 1MB). I heard that some vendors optimize their controllers for the default stripe size only. Is it true?
I have a several raids: RAID 0 SAS HDDs (each connected via 6Gb -- SAS-2). And 5+0 raid of SATA HDDs.
Both raids provide good read performance, but when I write to them, they start at insanely high speed (~1GB per second!). After some time, they become stuck for ~30 seconds. And then continue.
Write buffer is enabled.
Raids are from different vendors (Adaptec and IBM).
I believe they write data to the buffer and then need some time to flush it. It is reproducible only on huge (~ 20GB) files
Question 1: is it true? How can I profile/proof that fact?
I tried both: local access (copying from one raid to another and vice versa) and network access (via 10Gb network).
Effect is the same:
- 1GB/s
- After some time 0B/s
- After some time ~600MB/s
- And again.
Users are angry when copying process is frozen, so my question is does it work as it should? If so, what can we do except migrating to SDDs (30TB of SSDs are expensive) or limiting my network speed?
We spent some time improving our network performance (by enabling RSS, jumbo frames, migrating to latest SMB version that supports multiple threads etc) and we now have ~9.7Gb/s with iperf3 between client and server.
Question 2: Is it possible to have at least 300/400 megabytes per second for writing with 10 or 50 on HDDs? Shouldn't we have ~150MB/s for each disk, 300MB for 2 HDDs in RAID0? Big files utilize sequental access, aren't they?
Question 3: RAIDs have huge stripe size (about 1MB). I heard that some vendors optimize their controllers for the default stripe size only. Is it true?
Facebook
Twitter
Instagram