Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision 		Previous revision
hardware:computer [2015/07/30 17:03]
Jon Daniels 		hardware:computer [2024/02/14 17:55] (current)
Jon Daniels [Data Analysis]
Line 3: Line 3:
 ===== Acquisition ===== ===== Acquisition =====
  
-The main constraints is sufficiently fast disk write speed to handle the camera data.  Worst-case is 100 fps with full frame, or 800 MB/s (only one camera is used at a time).  Usually this is solved by using SSDs in RAID0 configuration (e.g. 4 SSDs in RAID0 can achieve >1 GB/s).  If you aren't using full frame or fastest imaging speed this requirement is relaxed.+Make sure to get a computer with sufficient PCI/PCIe slots for the camera framegrabber cards (usually 2 cameras/cards for dual-view) plus whatever other peripherals you need. 
 + 
 +Otherwise the main requirement having sufficiently fast disk write speed to handle the camera data.  Depending on the use case, solid state drives (SSDs) and/or RAID0 with SSDs may or may not be required.  Individual users should consider their requirements. 
 + 
 +The sCMOS cameras used with diSPIM can generate 800 MB/sec (100 fps at 4 MP, 16 bits per pixel).  However the maximum possible frame rate of the camera is not achieved for diSPIM.((Light sheet illumination only occurs during global exposure, and camera-limited frame rates occur without any global exposure time.))  Typical maximum acquisition speeds are 1024x1024 at 50fps or 512x512 at 200 fps; both these situations both generate 100MB/sec.  The average data rate, and hence hard drive speed requirement, is usually even less because most commonly acquisition occurs in bursts (i.e. there is time between successive time points) and a RAM buffer initially holds images so the hard drive needs to keep up with the average data rate.  Usually only one camera works at a time, though there are schemes where both cameras could be used simultaneously and thus double the data rate or else multiple cameras could be used for simultaneous multi-channel recording. 
 + 
 +100 MB/sec is typical for a magnetic hard drive.  300 MB/sec is typical for a single SSD.  If the data rate is too high for a single SSD, use SSDs in RAID0 configuration (e.g. 4 SSDs in RAID0 can achieve >1 GB/s).  Lately M.2 drives with PCIe interface with comparable speeds to a RAID0 with SSDs have become available and might be a good option.  To benchmark your PC's hard drive write speed you can use [[http://crystalmark.info/?lang=en | Crystal Disk Mark]].  I'm pretty sure the relevant score to diSPIM acquisition is the "Seq" "Write" score (Sequential (Block Size=1MiB) Read/Write with single Thread), at least for Micro-manager software with typical acquisition settings. 
 + 
 +Light sheet can generate lots of data very quickly, and it is important to have a plan to deal with the deluge.  This often involves support from the institution's IT department.  A helpful discussion of the challenges and options is the article [[https://arxiv.org/abs/2108.07631v1|Biologists need modern data infrastructure on campus]]. 
  
 ===== Data Analysis ===== ===== Data Analysis =====
  
-Having lots of RAM also helps the analysis so that the entire dataset can be held in active memory.  Ideally get a computer with CUDA-capable graphics card because some of the data analysis software can take advantage of it to speed the computation (OpenCL is a competing framework for GPU computation).  This is a nascent area and depends on software support; many developments in data analysis are forthcoming so it's hard to say exactly what will be the best in the long run.+Some users do image analysis and processing on a separate workstation, others use the acquisition computer when it's not being used for acquisition. 
 + 
 +Having lots of RAM can speed any analysis; ideally the entire dataset can be held in active memory.  Ideally get a computer with CUDA-capable graphics card because some of the data analysis software can take advantage of it to speed the computation (OpenCL is a competing framework for GPU computation).  This is a nascent area and depends on software support; many software developments data analysis are forthcoming so it's hard to say exactly what will be the best hardware in the long run. 
 + 
 +Micro-Manager 2.0 has a helpful ability to reslice data into the "lab frame" which is helpful especially for stage scanning data.  The GPU version of the algorithm can operate on datasets up to 1/4 of the GPU memory, e.g. 2 GB datasets can be processed on a GPU with 8 GB of working memory.
  
 ===== Specific suggestions ===== ===== Specific suggestions =====
  
-ASI has successfully used Dell Precision T3600 with 8-core Xeon CPU, 4x SSDs in RAID0, 64 GB RAM, Nvidia Quadro K4200, and enough PCIe slots (2 camera framegrabber cards, graphics card, and RAID controller).+In 2018 ASI has successfully used use a Dell Precision 7920 tower with 6-core Xeon CPU, one SSD for the OS/applications and a RAID0 drive with 4 SSDs for data, 64 GB RAM, and Nvidia P2000.  This has spare PCIe slots even with 2 camera framegrabber cards, the graphics card, and RAID controller. 
 + 
 +Previously (~2016?) ASI has successfully used Dell Precision T3600 with 8-core Xeon CPU, 4x SSDs in RAID0, 64 GB RAM, Nvidia Quadro K4200, and enough PCIe slots (2 camera framegrabber cards, graphics card, and RAID controller).
  

Trace: