Skip to end of metadata
Go to start of metadata

You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 64 Next »

How to Run daos on TACC Stampede2

TACC System information:

  • NO NVMe
  • NO Persistence Memory
  • Server will only use SCM (tmpfs which has size of ~90G only)

  • Node available with Queue Name

    Queue NameNode TypeMax Nodes per Job
    (assoc'd cores)*    		Max DurationMax Jobs in Queue*Charge Rate
    (per node-hour)
    skx-devSKX4 nodes
    (192 cores)*    		2 hrs1*1 SU
    skx-normalSKX128 nodes
    (6,144 cores)*    		48 hrs25*1 SU
    skx-large**SKX868 nodes
    (41,664 cores)*    		48 hrs3*1 SU

  • Stampede2 SKX Compute Node Specifications

    Model: Intel Xeon Platinum 8160 ("Skylake")
    Total cores per SKX node: 48 cores on two sockets (24 cores/socket)
    Hardware threads per core: 2
    Hardware threads per node: 48 x 2 = 96
    Clock rate: 2.1GHz nominal (1.4-3.7GHz depending on instruction set and number of active cores)
    RAM: 192GB (2.67GHz) DDR4
    Cache: 32KB L1 data cache per core; 1MB L2 per core; 33MB L3 per socket. Each socket can cache up to 57MB (sum of L2 and L3 capacity).
    Local storage: 144GB /tmp partition on a 200GB SSD. Size of /tmp partition as of 14 Nov 2017.

Server:

DAOS Server yaml 
# single server instance per config file for now
servers:
-
  targets: 16                		# Confirm the number of targets
  first_core: 0              		# offset of the first core for service xstreams
  nr_xs_helpers: 1           		# count of offload/helper xstreams per target
  fabric_iface: ib0          		# map to OFI_INTERFACE=ib0
  fabric_iface_port: 31416   		# map to OFI_PORT=31416
  log_mask: ERR     		 		# map to D_LOG_MASK=ERR
  log_file: /tmp/daos_server.log 	# map to D_LOG_FILE=/tmp/server.log

  # Environment variable values should be supplied without encapsulating quotes.
  env_vars:                 # influence DAOS IO Server behaviour by setting env variables
  - CRT_TIMEOUT=120
  - CRT_CREDIT_EP_CTX=0
  - PSM2_MULTI_EP=1
  - CRT_CTX_SHARE_ADDR=1
  - PMEMOBJ_CONF=prefault.at_open=1;prefault.at_create=1;  # Do we need this?
  - PMEM_IS_PMEM_FORCE=1								   # Do we need this?

  # Storage definitions

  # When scm_class is set to ram, tmpfs will be used to emulate SCM.
  # The size of ram is specified by scm_size in GB units.
  scm_mount: /dev/shm   # map to -s /mnt/daos
  scm_class: ram
  scm_size: 90

  # When scm_class is set to dcpm, scm_list is the list of device paths for
  # AppDirect pmem namespaces (currently only one per server supported).
  # scm_class: dcpm
  # scm_list: [/dev/pmem0]

  # If using NVMe SSD (will write /mnt/daos/daos_nvme.conf and start I/O
  # service with -n <path>)
  # bdev_class: nvme
  # bdev_list: ["0000:81:00.0"]  # generate regular nvme.conf

  # If emulating NVMe SSD with malloc devices
  # bdev_class: malloc  # map to VOS_BDEV_CLASS=MALLOC
  # bdev_size: 4                # malloc size of each device in GB.
  # bdev_number: 1              # generate nvme.conf as follows:
              # [Malloc]
              #   NumberOfLuns 1
              #   LunSizeInMB 4000

  # If emulating NVMe SSD over kernel block device
  # bdev_class: kdev            # map to VOS_BDEV_CLASS=AIO
  # bdev_list: [/dev/sdc]       # generate nvme.conf as follows:
              # [AIO]
              #   AIO /dev/sdc AIO2

  # If emulating NVMe SSD with backend file
  # bdev_class: file            # map to VOS_BDEV_CLASS=AIO
  # bdev_size: 16           # file size in GB. Create file if does not exist.
  # bdev_list: [/tmp/daos-bdev] # generate nvme.conf as follows:
              # [AIO]
              #   AIO /tmp/aiofile AIO1 4096

Server Environment variables (If set any)

Client Configuration:

Configuration:

Environment variables (If set any):

Run time Environment 
export CRT_PHY_ADDR_STR="ofi+psm2"
export OFI_INTERFACE=ib0
export FI_PSM2_NAME_SERVER=1
export PSM2_MULTI_EP=1
export FI_SOCKETS_MAX_CONN_RETRY=1
export CRT_CTX_SHARE_ADDR=1
export CRT_TIMEOUT=120

Other important information for running test:

Item or NotesDescription
For Defect CART-777There is an issue with verbs+open MPI so some time it orterun is getting stuck because of it, To workaround the issue use the "--mca btl tcp,self --mca oob tcp" with IOR or other commands



Test Description:

Testing Area Test

Test Priority (1- HIGH,  2 - LOW)

Number of ServersNumber of ClientsInput ParameterExpected ResultObserved ResultDefectNotesExpected SU's (1 node * 1 hour = 1 SU)
Server YAML config options To verify the test cases from below section with specific server config options in YAML file 1



target = [16]

nr_xs_helpers = [1]

CRT_CTX_SHARE_ADDR=[0, 1]

No sever crash,

Performance increase linearly 



No need individual test but below test can be used this configuration 

Performance

No Replica 

Run IOR and collect BW

Run IOR small size and collect IOPS

1

1,

8,

32,

128

128

1,

16,

96,

256

740

protocol : daos

Transfer Size: 256B 4K 128K 512K 1M (Do we need non standard size also be covered?)

Block Size: 64M (Depend upon no. of process as file size will increase because of it)

FPP and SSF

single server got ~12GB Read/write so it should scale linearly.

With 128 server should be close to 1.5TB  BW? 




1406 Nodes

taking ~30 min 


703

Replica 2 Way

Run IOR and collect BW

Run IOR small size and collect IOPS

1

8,

32,

128

16,

96,

740

Same As Above



1020 Nodes

for ~30 min

510

Replica 3 Way

Run IOR and collect BW

Run IOR small size and collect IOPS

1

8,

32,

128

16,

96,

740

Same As Above



1020 Nodes

for ~30 min

510

Replica 4 Way

Run IOR and collect BW

Run IOR small size and collect IOPS

1

8,

32,

128

16,

96,

740

Same As Above



1020 Nodes

for ~30 min

510

Any Erasure Encoding object class need to  run? May be with medium size?

EC_2P1G1
EC_2P2G1
EC_8P2G1

1?3296Same As Above?



128 nodes for ~60 min120
Metadata Test (Using MDTest)1

1,

8,

32,

128

128

1,

16,

96,

256

740

How many tasks per client 1 ,4 or only 8?

What class type should be tested ? 

-n = 1000 (every process will creat/stat/read/remove )

-z = 0 and 20 (depth of hierarchical directory structure) 

Result with 1 server, 1 client is available from

https://jira.hpdd.intel.com/secure/attachment/31383/sbatch_run.txt




1406 Nodes

taking ~15 min 


350
CART self_test1

2

32

126

1

1

1

orterun --timeout 3600 --mca mtl ^psm2,ofi -x FI_PSM2_DISCONNECT=1 -np 1 -ompi-server <urifile> self_test --group-name daos_server --endpoint 0-<NO_OF_SERVER>:0 --master-endpoint 0-<NO_OF_SERVER>:0 --message-sizes 'b1048576',' b1048576 0','0 b1048576',' b1048576 i2048',' i2048 b1048576',' i2048',' i2048 0','0 i2048','0' --max-inflight-rpcs 1 --repetitions 100
Did not get all the number for 126 servers CART-791


https://wiki.hpdd.intel.com/download/attachments/114950812/2SN_1CN_TACC-Stampede2_20191022_144511.txt?api=v2


https://wiki.hpdd.intel.com/download/attachments/114950812/32SN_1CN_TACC-Stampede2_20191022_172517.txt?api=v2


https://wiki.hpdd.intel.com/download/attachments/114950812/126SN_1CN_TACC-Stampede2_20191023_091546.txt?api=v2

166 Nodes

for ~5 min

14
POSIX (Fuse)2?3296Run IOR with POSIX mode. Are we there to get the full performance ? 



128

for ~60 min

128
DFS2

Not sure if we want to cover dfs as we are covering daos with IOR on above test cases





HDF5?2?3296Any specific test we want to run?





FIO?


Do we want to test this?





Functionality and Scale testingRun all daos_test2128740




868 node for ~60 min

Single server/Max clients

(IOR)


1

126

(Client processes

1

64

128

512

1024

2016)

Create pool, Query pool

Run IOR (Specific size?)

Transfer size: 256B, 1M

Block size: 16M for 256B TS, otherwise 64M

Flags: -w -W -r -R

iter: 3

Poole create should work fine. IOR will run with ~2000 tasks so it should success. Query pool info after IOR run and measure the pool size compare to file size.

Assuming 16 client processes per node. (Need to verify if it works fine. 8 client processes per node works.)



(Total nodes available at present 127. 16 Client processes per node)128 node for ~30 min64

1

866

(Client processes

1

128

1024

4096

8192

13856)

Create pool, Query pool

Run IOR (Specific size?)

Transfer size: 256B, 1M

Block size: 16M for 256B TS, otherwise 64M

Flags: -w -W -r -R

iter: 3

Poole create should work fine. IOR will run with ~13000 tasks so it should success. Query pool info after IOR run and measure the pool size compare to file size.

Assuming 16 client processes per node. (Need to verify if it works fine. 8 client processes per node works.)



(Total nodes available at present 867. 16 Client processes per node)868 node for ~30 min434

Max servers/single client

(IOR)


1

8

16

32

64

126

1

(Client processes

16)

Create pool, Query pool

Run IOR with DAOS and POSIX api

Transfer size: 256B, 1M

Block size: 16M for 256B TS, otherwise 64M

Flags: -w -W -r -R

iter: 3

Poole create should work fine. IOR will be run with 16 client processes per node (need to verify if it works fine. 8 client processes per node works). Query pool info after IOR run and measure the pool size compare to file size.

(Total nodes available at present 127. 16 Client processes per node)128 node for ~30 min64

1

16

64

256

512

866

1

(Client processes

16)

Create pool, Query pool

Run IOR with DAOS and POSIX api

Transfer size: 256B, 1M

Block size: 16M for 256B TS, otherwise 64M

Flags: -w -W -r -R

iter: 3

Poole create should work fine. IOR will be run with 16 client processes per node (need to verify if it works fine. 8 client processes per node works). Query pool info after IOR run and measure the pool size compare to file size.

(Total nodes available at present 867. 16 Client processes per node)868 node for ~30 min434

Single server/Max clients

(Mdtest)


1

Client processes

1

64

128

512

1024

2032

Create pool

Run mdtest with DFS and Posix api

num of files/dir: -n 100|10K

write and read (-w and -e): 4 for -n 100 otherwise keep files empty

depth (-z): 0 and 20

iter: 3

Pool create should work as expected. Run mdtest with 16 client processes per node(need to be verified if 16 clients per node works fine).

(Total nodes available at present 127. 16 Client processes per node)128 node for ~30 min64

1

Client processes

1

128

1024

4096

8192

13872

Create pool

Run mdtest with DFS and Posix api

num of files/dir: -n 100|10K

write and read (-w and -e): 4 for -n 100 otherwise keep files empty

depth (-z): 0 and 20

iter: 3

Pool create should work as expected. Run mdtest with 16 client processes per node(need to be verified if 16 clients per node works fine).

(Total nodes available at present 867. 16 Client processes per node)868 node for ~30 min434

Max servers/single client

(Mdtest)


1

8

16

32

64

127

Client processes

16

Create pool

Run mdtest with DFS and Posix api

num of files/dir: -n 100|10K

write and read (-w and -e): 4 for -n 100 otherwise keep files empty

depth (-z): 0 and 20

iter: 3

Pool create should work as expected. Run mdtest with 16 client processes per node(need to be verified if 16 clients per node works fine).

(Total nodes available at present 127. 16 Client processes per node)128 node for ~30 min64

1

16

64

256

512

867

Client processes

16

Create pool

Run mdtest with DFS and Posix api

num of files/dir: -n 100|10K

write and read (-w and -e): 4 for -n 100 otherwise keep files empty

depth (-z): 0 and 20

iter: 3

Pool create should work as expected. Run mdtest with 16 client processes per node(need to be verified if 16 clients per node works fine).

(Total nodes available at present 867. 16 Client processes per node)868 node for ~30 min434

Large number of Pools (~1000)



128

Server number  seems ok?

740

Create large number of pools (~90MB each),

Write small data with IOR.

Restart all the servers.

Query all the pools

Read the IOR data from each pool with verification

what other operation needed after pool creation? 


Measure server restart time with this many pools

Pool query should report correct sizes after IOR write

IOR read should work fine with data validation after all server restart




868 node for ~60 min 868

dmg utility testing

for example: pool query




dmg pool create 

dmg pool query

dmg pool destroy

Anything more to cover? Some of this tools are going to cover in other test cases







Negative Scenarios with ScalabilityServer failure and rebuild data1287401

Create the multiple pools.

Store the IOR with 2,3,4 replica and with multiple groups.

Kill server one by one 64 maximum (Half the requested size)?

After each server kill read the IOR data and verify the content.

Multiple server can be killed (2/4/8), Object data will be lost if all copy lost. May be we can verify the remaining system is functional

Rebuild should happen for all the object and data should not be corrupted after server failure


868 for ~2 hours868
daos_run_io_conf1287402

This will exclude the ranks and add it back in to the loop for given number.

We can have maximum 16 targets and include all rank. Test will exclude the rank randomly and add it back.

Pool query is also part of this test to verify the usage 

We have not tried on TACC but locally it works but there are few issue need to be resolved which we caught during local testing (DAOS-3510)


868 for ~30 min434
Reliability and Data Integrity (Soak testing)Current Soak testing







868

for 2 hours 

1736






















Question:

  • Do we need to test other replica other than what mention in above?
  • Any Erasure encoding class needs to be tested ?

Estimation:

Test Plan versionEstimated SU
v. 477605




  • No labels