ALPACA I/O rates

19apr19

links
Operational parameters
I/O rates

**operational parameters (updated 06may19)**
Number of beams	40
number of pols	2	full stokes available,a,b,axb,bxa
frequency range	1289-1720Mhz
total bandwidth	440 MHz
input bandwidth to beam former	300 MHz	selectable within the 440 MHz band
course channels per pol	683
course channel sep/bandwidth	439.5KHz sep	595.9 Khz Bw
fine individual channel bw	4.6Khz	32784 chan
fine channel individual bw	150 MHz
sample rate (course)	64 usecs
sample rate (fine)	> .1 secs

**BeamFormer output I/O Rates**
smpTime	BW MHz	#freqChan per pol	#pols	output bits	I/O rate Mbyte/Bm	Total I/O rate (all bms ) MByte/sec
course resolution
64usecs	300	683	4(stokes)	32	171	6830
			4(stokes)	16	85.5	3415
			2	32	85.5	3415.
			2	16	42.75	1707
			1	32	42.75	1707
			1	16	21.375	853.5
			1	8	10.688	426.75
			1	4	5.344	213.375
fine resolution
.1 sec	150	32784	4(stokes)	32	5.26	210.4
.1 sec	150	32784	2	32	2.63	105.2

Component i/o rates

FPGA BRDS:

18 Boards
138 (2pol*69dipoles) dipole input streams
8 dipoles per fpga board
fpga output rate for each board:

683 pfb channels, 585.9 KHz per channel output each dipole channel, two bytes per output (8bit real, 8bit img)
585.9Khz * 2bytes/smp * 683Chan * 8dipoles/fpgaBrd = 6.40272Gbytes/sec = 51.2217 Gbits/secs

GPU/HPC inputs

50 gpus
2 gpus/hpc
total # freqchannels*pols

683freqChan *2pols=1366 chan

freqChanPols per gpu

1366/50 = 27.32--> 28 freqchanpol/gpu (14freqchan *2pol)

ioRate 1freqChan

585.9Khz*2bytes/smp * 69 dipoles = 80.8542 MBytes/sec

Input ioRate 1 gpu

80.8542 MB/sec *28 freqChanPol/gpu = 2.263918 GBytes/sec

input ioRate 1 hpc:

2gpu/hpc : 2.26391 GBytes/sec*2 = 4.5278354GBytes/sec => 36.22 Gbits/sec

Note:

Once the 69 dipoles for a single freq channel are input to a gpu, it can compute all 40 beams, by using different sets of weights (one for each beam)

GPU/HPC outputs

output: spectral density, polA,polB,BxA and AxB 4 byte floats
min integration time pulsar mode: 64e-6
beams each freq chan: 40
iorate 1 freq chan_40bms: 1./64e-6 * 4bytes * 4pols * 40beams=10MBytes/sec
freqChan/Gpu: 28 (14*2pols)
output rate 1gpu: 10Mbytes/sec * 14 freqchan=140MBytes/sec
output rate 1 hpc: 2gpus = 280MBytes/sec
output rate all: 683 freqchan* 10Mybtes/sec (freqchan40bms) = 6.830GBytes/sec -> 54.64 Gbit/sec

GPU/HPC outputs beam formed voltages.

freqChanWidth: 585.9Khz
number of freqChan: 683 chan * 2 pols
freqchanPols/gpu: 28 (14*2pols)
output complex floats 2*4bytes per sample
total output i/o rate:

per gpu: 14chan*2pol*40bms*585.9Khz * 8bytes/sample=5.249 GBytes/sec
per hpc: 2*gpu = 10.4993 GBytes/sec
total: 683Chan*2pol*40bms*585.9Khz*8bytes/sample=260.480GBytes/sec = 2.084 Tbits/sec

<- page up
home_~phil