I would like to share with you the practical experience of testing and using the HPE Nimble storage system, in our case, we will share our impressions regarding the All-Flash model - HPE Nimble AF40. How relevant and interesting this article will be is up to you, but at one time we were faced with the problem of finding real stories (do not confuse with success stories), in which the author would mostly share his thoughts, experiences and impressions, i.e. pure emotions from the heart, not dry facts or ubiquitous marketing. Therefore, we decided to spend a little of our time and try to write something like that, so do not judge strictly, because this is the first try of the pen. So let's go.
In order to introduce you a little into the course of the matter, for a start, we will give scanty information about our experimental. This model is positioned by HPE among its own family of Nimble All-Flash arrays as the best choice in terms of price / performance ratio. In terms of performance, the system is quite different from the younger AF20 model (almost 5 times more!), But at the same time it is not so much inferior to the older models - AF60 and AF80. The oldest model, our copy is on average 3 times inferior in the passport, even less in some indicators, and the Nimble AF60 model is no more than 2 times. The arrays of the entire HPE Nimble line have all the necessary functionality, according to their status - arrays of the middle price segment. Here you have both compression with deduplication, and the latter with a variable block, and hardware snapshots with agreement at the application level,
Now about the numbers
According to the passport, the system allows you to get up to 130K IOPS on 4K blocks in a mixed load of 50/50 on random I / O operations, 150K IOPS on all the same 4K blocks at 100% random read and 150K IOPS at 100% random write - all this without taking into account deduplication. Taking it into account, the final indicators will depend on its overall coefficient. For example, with an overall deduplication ratio of 10: 1, random write operations will sag slightly - up to 130K on 4K blocks. At the same time, the number of IOPS on read operations will increase to 170K, and in a mixed mode 50/50, 96K IOPS are declared. By today's standards - more than modest for an All-Flash array. Moreover, the array is a representative of the Mid-Range segment, and you expect really decent numbers from them, since the segment is a kind of a racing track, where every fraction of a second counts and manufacturers put up the best, just to amaze the sophisticated audience and win their hearts (and at the same time their wallets). There is nothing to amaze here in terms of numbers, especially since now even Entry-Level (entry-level) arrays can boast of 300 + K IOPS, which, by the way, could be the envy of a decent number of Hi-End data storage systems 8-10 years ago. Yes, progress does not stand still - everything is so. Another question is that the selection criteria are different for everyone and not everyone needs the fastest car. I am so much more of an adherent of comfort. Yes, dynamics, undoubtedly, is an important and necessary thing (to a certain extent) for the formation of that very comfort: when the car is driving and allows without tension, when it is necessary to add, to make a dynamic maneuver and in general it is pleasant when a powerful engine is under the hood, which, perhaps, will never be used in full force, but who cares ... Sports cars that cheerfully and cheerfully rumble in the city from traffic lights to traffic lights, notably shaking the carcasses of their remote owners - very dynamic acceleration and no less dynamic braking. The taste and color of felt-tip pens are different: some like it, some don't. I am so with all hands for practicality, i.e. you can eat soup with a fork, but why, when there is a spoon. Or why eat borsch with a scoop from a 10-liter saucepan, if there is a plate and the same spoon. Well, in general, I think you understand me. notably shaking the carcasses of their daring owners - very dynamic acceleration and no less dynamic braking. The taste and color of felt-tip pens are different: some like it, some don't. I am so with all hands for practicality, i.e. you can eat soup with a fork, but why, when there is a spoon. Or why eat borsch with a scoop from a 10-liter saucepan, if there is a plate and the same spoon. Well, in general, I think you understand me. notably shaking the carcasses of their daring owners - very dynamic acceleration and no less dynamic braking. The taste and color of felt-tip pens are different: some like it, some don't. I am so with all hands for practicality, i.e. you can eat soup with a fork, but why, when there is a spoon. Or why eat borsch with a scoop from a 10-liter saucepan, if there is a plate and the same spoon. Well, in general, I think you understand me.
. , . , . , , . , : , .. ( , ).
, HPE - , . , – . , , … (, – , ).
, , : , , best practice . , , ( ), – End of General Support. , SAN- SN6000 (8- Qloqic SANbox 5800 — 2010- ). , 4 , — . 16 FC HPE Nimble AF40 , , . , , , best practice, .. 4 - , 2 16 FC 8 .
DL385 Gen7 VMware ESXi 6.0.0. VMware vCenter appliance Linux. AMD Opteron 6180 SE 192 DDR3 PC3-10600 ECC RDIMM (1300 ). SAN- FC HBA 8 (HPE 81Q) . HPE Nimble AF40 24 480 , 16 FC HBA .
— vdbench, - HCIbench , VMware.
HPE Nimble AF40 , . , , , , . : - . .
№1: ,
, . , , : , , ( ). – . . 4U, , «» — 89 , — 50 . , — «». , , . — , , . 10 — , , . , , ; .
. Welcome Center Nimble. - -, , . Nimble , — HPE Primera.
, , — -, , , .
, , , Installation Guide:
15-30 . , , — «5» «5». , , , .
№2: HPE Nimble AF40 VMware
c HPE Nimble — VMware VVOL. , VVOL, VMware VMFS Datastores . , LUN ( VMware VMFS Datastore , , ).
QoS . , , VMware VMFS Datastore. . , VMware VMFS Datastore - , . , , — , .
, , , . , , VMware VMFS Datastore, . , , — ESXi 6.0.0 256. , .
VVOL . , VMware vCenter. , , - HPE Nimble AF40 . ( , , ). vCenter VM Storage Policies.
( SAN , , ).
1. vCenter
2. HPE Nimble – Folder VVOLs:
3. Performance Policies HPE Nimble VM Storage Policies ( Performance Policies HPE Nimble — , , ):
4. VM Storage Policies vCenter Performance Policies:
5. Folder vCenter:
VVOL HPE Nimble vCenter, VM Storage Policies:
№3:
. — 100- 4 ( ). HCIbench :
— 12, — 8 10 , — 6 , — 1 , — 384. — , 0, 1:1. , .. .
. HPE Nimble AF40:
HCIbench ( : GMT -07:00, 10 GMT +03:00):
— 2 , — : , 6, 192:
.
, , QoS, VVOL 20 IOPS 384 , — 10 .
SSH- ESXi esxtop, , 125-128 FC HBA . : QoS, - 20 000, -.
, . , .
? 100% , – 4, – . – 384 .
, 6 , , , 192 .
2 . 100% . , IOPS, , . , 150 IOPS 100% 4 100% 4. , , HPE Nimble CASL , .
, , vdbench , , , , , ( , ).
: , 16 FC, , , - .
, SkyNet HPE Infosight, , .
№4: HPE Infosight – ,
HPE Infosight , . , . , – . , .
HPE Infosight ( , ), , .
. , Operational, , Recommendations, . , , , c HPE Nimble ( HPE Nimble Cross Stack Recommendations).
— Physical CPU Saturation . — Elevated Latency, Virtual CPU Overprovisioning, :
(Executive), , , , ( ), . Wellness , , Capacity .
, , — HPE Infosight, .
Cross Stack Recommendations, Resource Planner, - , . .
, , , , HP Virtualization Performance Viewer, , , , .
, , AI Performance Recommendations . , — , .
, , — ( , , Excel- ), , , .
: « , - - VMware vCenter, - — HPE Nimble Monitoring, ». , , . Infrastructure -> Virtualization , ( VMware Hyper-V) -, , , - ( , , ).
, , .
HPE Nimble. , , , ( ) , .
: HPE Infosight , , – . 86% . ? “” , , , . , , “” . , , . HPE Nimble, – ( ?) – 3 , .
№5: ,
, — ! - , , , , “”. , , , . , . , , , . , , – ?
, , - . , . L3 HPE Nimble. (HA Failover). , . , , , . , HPE Infosight L3 HPE Nimble, . , , , , ( , , – ). , , . .
, , ( ). , , , - -. - HPE Nimble . , . . , , .. 2-3 , , , . HPE Nimble (, ).
, , , , , HPE Nimble, .. . , .
, : , - ? , . , , , . – . , 20% SSD/80% MDL 7,2K, , HPE Nimble HF20 HPE Nimble HF40, , . , .. , , . , HPE Infosight, , - , - , , , , – . VDI? , ! VMware Horizon Just-in-Time (linked clones ). - NVIDIA Tesla T4 RTX6000. .
UPD: within the framework of the article, I missed a lot of interesting things, namely, the primary problem of testing All-Flash arrays on VMware ESXi hypervisors, with which they fought for a long time and sinned on HPE Nimble (we did not say that we are saints). Out of the box, load testing will not work adequately: there are problems with queues on FC HBAs (driver base values) - DQLEN and No of outstanding IOs with competing worlds, as well as restrictions on the part of virtual controllers of virtual machines. This material will be quite enough for a separate article.