Block Storage

We are currently running a VSP 5500-6N NVMe System where we started compression&dedup (post) on ESX ldevs as well as compression on some development databases. As expected the overall CPU usage raised from about 30% (before we started using ADR) to about 55-60% CPU usage. Our overall workload within the last 150days average is 37% write (91'000), 63% read (156'000) IOs and 28% write (3,6GB/s), 72% read (9.2 GB/s) throughput.

On the ESX ldevs (compression&dedup) we around 4:1, the DB ldevs (compression) 2:1. 

We do have around 1 PB of data where we would like to activate compression as well, but due to the overall CPU usage of 55-60% we got the idea to upgrade the system from 5500 to a 5600 with compression cards. We hope that this would lower our overall CPU usage as the compression would be done on the compression cards and no longer on the CPU. In addition, the 5600 should have more powerful processors what should also have a positive impact on CPU usage (as we were told that CPU usage of 60-70% might have  negative impact on response time, over 80% CPU usage it gets worse)

The questions now are: 

• if someone did an upgrade and what was the impact on overall CPU usage
• if someone already did an upgrade from 5500 to 5600 with compression cards and what the experience was
• if we test the 5600 with ADR and we would not get the expected results would it be possible to downgrade from 5600 with compression cards back to the 5500 ? Any experiences ?

Any feedback would be appreciated.

------------------------------
Patrick Ammann 2020
PostFinance AG
------------------------------

Hi @Patrick Ammann 2020 - There's an internal sizing tool called CPK that helps model how much IOPS/throughput a certain storage model & configuration can deliver. It requires random I/O block size, sequential I/O block size, read %, random %, cache hits %, and storage model & config. If you can provide these details to your HV rep, he/she can model the max performance for your workload on 5500 vs 5600.

We are currently running a VSP 5500-6N NVMe System where we started compression&dedup (post) on ESX ldevs as well as compression on some development databases. As expected the overall CPU usage raised from about 30% (before we started using ADR) to about 55-60% CPU usage. Our overall workload within the last 150days average is 37% write (91'000), 63% read (156'000) IOs and 28% write (3,6GB/s), 72% read (9.2 GB/s) throughput.

On the ESX ldevs (compression&dedup) we around 4:1, the DB ldevs (compression) 2:1. 

We do have around 1 PB of data where we would like to activate compression as well, but due to the overall CPU usage of 55-60% we got the idea to upgrade the system from 5500 to a 5600 with compression cards. We hope that this would lower our overall CPU usage as the compression would be done on the compression cards and no longer on the CPU. In addition, the 5600 should have more powerful processors what should also have a positive impact on CPU usage (as we were told that CPU usage of 60-70% might have  negative impact on response time, over 80% CPU usage it gets worse)

The questions now are: 

• if someone did an upgrade and what was the impact on overall CPU usage
• if someone already did an upgrade from 5500 to 5600 with compression cards and what the experience was
• if we test the 5600 with ADR and we would not get the expected results would it be possible to downgrade from 5600 with compression cards back to the 5500 ? Any experiences ?

Any feedback would be appreciated.

------------------------------
Patrick Ammann 2020
PostFinance AG
------------------------------

HI @Dang Luong - i'm not that much interested in max performance numbers. I would be more interested in comparing the two controlles side by side with the same configuration (same amount of IOPS/throuput, same IO pattern) and get an ideo of how much more CPU would be available within the 5600 controller that could be used for compression and/or compression and deduplication. 

Hi @Patrick Ammann 2020 - There's an internal sizing tool called CPK that helps model how much IOPS/throughput a certain storage model & configuration can deliver. It requires random I/O block size, sequential I/O block size, read %, random %, cache hits %, and storage model & config. If you can provide these details to your HV rep, he/she can model the max performance for your workload on 5500 vs 5600.

We are currently running a VSP 5500-6N NVMe System where we started compression&dedup (post) on ESX ldevs as well as compression on some development databases. As expected the overall CPU usage raised from about 30% (before we started using ADR) to about 55-60% CPU usage. Our overall workload within the last 150days average is 37% write (91'000), 63% read (156'000) IOs and 28% write (3,6GB/s), 72% read (9.2 GB/s) throughput.

On the ESX ldevs (compression&dedup) we around 4:1, the DB ldevs (compression) 2:1. 

We do have around 1 PB of data where we would like to activate compression as well, but due to the overall CPU usage of 55-60% we got the idea to upgrade the system from 5500 to a 5600 with compression cards. We hope that this would lower our overall CPU usage as the compression would be done on the compression cards and no longer on the CPU. In addition, the 5600 should have more powerful processors what should also have a positive impact on CPU usage (as we were told that CPU usage of 60-70% might have  negative impact on response time, over 80% CPU usage it gets worse)

The questions now are: 

• if someone did an upgrade and what was the impact on overall CPU usage
• if someone already did an upgrade from 5500 to 5600 with compression cards and what the experience was
• if we test the 5600 with ADR and we would not get the expected results would it be possible to downgrade from 5600 with compression cards back to the 5500 ? Any experiences ?

Any feedback would be appreciated.

------------------------------
Patrick Ammann 2020
PostFinance AG
------------------------------